diff --git a/llm/ant/ant.go b/llm/ant/ant.go
new file mode 100644
index 0000000..dce17f1
--- /dev/null
+++ b/llm/ant/ant.go
@@ -0,0 +1,480 @@
+package ant
+
+import (
+	"bytes"
+	"cmp"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"math/rand/v2"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	"sketch.dev/llm"
+)
+
+const (
+	DefaultModel = Claude37Sonnet
+	// See https://docs.anthropic.com/en/docs/about-claude/models/all-models for
+	// current maximums. There's currently a flag to enable 128k output (output-128k-2025-02-19)
+	DefaultMaxTokens = 8192
+	DefaultURL       = "https://api.anthropic.com/v1/messages"
+)
+
+const (
+	Claude35Sonnet = "claude-3-5-sonnet-20241022"
+	Claude35Haiku  = "claude-3-5-haiku-20241022"
+	Claude37Sonnet = "claude-3-7-sonnet-20250219"
+)
+
+// Service provides Claude completions.
+// Fields should not be altered concurrently with calling any method on Service.
+type Service struct {
+	HTTPC     *http.Client // defaults to http.DefaultClient if nil
+	URL       string       // defaults to DefaultURL if empty
+	APIKey    string       // must be non-empty
+	Model     string       // defaults to DefaultModel if empty
+	MaxTokens int          // defaults to DefaultMaxTokens if zero
+}
+
+var _ llm.Service = (*Service)(nil)
+
+type content struct {
+	// TODO: image support?
+	// https://docs.anthropic.com/en/api/messages
+	ID   string `json:"id,omitempty"`
+	Type string `json:"type,omitempty"`
+	Text string `json:"text,omitempty"`
+
+	// for thinking
+	Thinking  string `json:"thinking,omitempty"`
+	Data      string `json:"data,omitempty"`      // for redacted_thinking
+	Signature string `json:"signature,omitempty"` // for thinking
+
+	// for tool_use
+	ToolName  string          `json:"name,omitempty"`
+	ToolInput json.RawMessage `json:"input,omitempty"`
+
+	// for tool_result
+	ToolUseID  string `json:"tool_use_id,omitempty"`
+	ToolError  bool   `json:"is_error,omitempty"`
+	ToolResult string `json:"content,omitempty"`
+
+	// timing information for tool_result; not sent to Claude
+	StartTime *time.Time `json:"-"`
+	EndTime   *time.Time `json:"-"`
+
+	CacheControl json.RawMessage `json:"cache_control,omitempty"`
+}
+
+// message represents a message in the conversation.
+type message struct {
+	Role    string    `json:"role"`
+	Content []content `json:"content"`
+	ToolUse *toolUse  `json:"tool_use,omitempty"` // use to control whether/which tool to use
+}
+
+// toolUse represents a tool use in the message content.
+type toolUse struct {
+	ID   string `json:"id"`
+	Name string `json:"name"`
+}
+
+// tool represents a tool available to Claude.
+type tool struct {
+	Name string `json:"name"`
+	// Type is used by the text editor tool; see
+	// https://docs.anthropic.com/en/docs/build-with-claude/tool-use/text-editor-tool
+	Type        string          `json:"type,omitempty"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema,omitempty"`
+}
+
+// usage represents the billing and rate-limit usage.
+type usage struct {
+	InputTokens              uint64  `json:"input_tokens"`
+	CacheCreationInputTokens uint64  `json:"cache_creation_input_tokens"`
+	CacheReadInputTokens     uint64  `json:"cache_read_input_tokens"`
+	OutputTokens             uint64  `json:"output_tokens"`
+	CostUSD                  float64 `json:"cost_usd"`
+}
+
+func (u *usage) Add(other usage) {
+	u.InputTokens += other.InputTokens
+	u.CacheCreationInputTokens += other.CacheCreationInputTokens
+	u.CacheReadInputTokens += other.CacheReadInputTokens
+	u.OutputTokens += other.OutputTokens
+	u.CostUSD += other.CostUSD
+}
+
+type errorResponse struct {
+	Type    string `json:"type"`
+	Message string `json:"message"`
+}
+
+// response represents the response from the message API.
+type response struct {
+	ID           string    `json:"id"`
+	Type         string    `json:"type"`
+	Role         string    `json:"role"`
+	Model        string    `json:"model"`
+	Content      []content `json:"content"`
+	StopReason   string    `json:"stop_reason"`
+	StopSequence *string   `json:"stop_sequence,omitempty"`
+	Usage        usage     `json:"usage"`
+}
+
+type toolChoice struct {
+	Type string `json:"type"`
+	Name string `json:"name,omitempty"`
+}
+
+// https://docs.anthropic.com/en/api/messages#body-system
+type systemContent struct {
+	Text         string          `json:"text,omitempty"`
+	Type         string          `json:"type,omitempty"`
+	CacheControl json.RawMessage `json:"cache_control,omitempty"`
+}
+
+// request represents the request payload for creating a message.
+type request struct {
+	Model         string          `json:"model"`
+	Messages      []message       `json:"messages"`
+	ToolChoice    *toolChoice     `json:"tool_choice,omitempty"`
+	MaxTokens     int             `json:"max_tokens"`
+	Tools         []*tool         `json:"tools,omitempty"`
+	Stream        bool            `json:"stream,omitempty"`
+	System        []systemContent `json:"system,omitempty"`
+	Temperature   float64         `json:"temperature,omitempty"`
+	TopK          int             `json:"top_k,omitempty"`
+	TopP          float64         `json:"top_p,omitempty"`
+	StopSequences []string        `json:"stop_sequences,omitempty"`
+
+	TokenEfficientToolUse bool `json:"-"` // DO NOT USE, broken on Anthropic's side as of 2025-02-28
+}
+
+const dumpText = false // debugging toggle to see raw communications with Claude
+
+func mapped[Slice ~[]E, E, T any](s Slice, f func(E) T) []T {
+	out := make([]T, len(s))
+	for i, v := range s {
+		out[i] = f(v)
+	}
+	return out
+}
+
+func inverted[K, V cmp.Ordered](m map[K]V) map[V]K {
+	inv := make(map[V]K)
+	for k, v := range m {
+		if _, ok := inv[v]; ok {
+			panic(fmt.Errorf("inverted map has multiple keys for value %v", v))
+		}
+		inv[v] = k
+	}
+	return inv
+}
+
+var (
+	fromLLMRole = map[llm.MessageRole]string{
+		llm.MessageRoleAssistant: "assistant",
+		llm.MessageRoleUser:      "user",
+	}
+	toLLMRole = inverted(fromLLMRole)
+
+	fromLLMContentType = map[llm.ContentType]string{
+		llm.ContentTypeText:             "text",
+		llm.ContentTypeThinking:         "thinking",
+		llm.ContentTypeRedactedThinking: "redacted_thinking",
+		llm.ContentTypeToolUse:          "tool_use",
+		llm.ContentTypeToolResult:       "tool_result",
+	}
+	toLLMContentType = inverted(fromLLMContentType)
+
+	fromLLMToolChoiceType = map[llm.ToolChoiceType]string{
+		llm.ToolChoiceTypeAuto: "auto",
+		llm.ToolChoiceTypeAny:  "any",
+		llm.ToolChoiceTypeNone: "none",
+		llm.ToolChoiceTypeTool: "tool",
+	}
+
+	toLLMStopReason = map[string]llm.StopReason{
+		"stop_sequence": llm.StopReasonStopSequence,
+		"max_tokens":    llm.StopReasonMaxTokens,
+		"end_turn":      llm.StopReasonEndTurn,
+		"tool_use":      llm.StopReasonToolUse,
+	}
+)
+
+func fromLLMCache(c bool) json.RawMessage {
+	if !c {
+		return nil
+	}
+	return json.RawMessage(`{"type":"ephemeral"}`)
+}
+
+func fromLLMContent(c llm.Content) content {
+	return content{
+		ID:           c.ID,
+		Type:         fromLLMContentType[c.Type],
+		Text:         c.Text,
+		Thinking:     c.Thinking,
+		Data:         c.Data,
+		Signature:    c.Signature,
+		ToolName:     c.ToolName,
+		ToolInput:    c.ToolInput,
+		ToolUseID:    c.ToolUseID,
+		ToolError:    c.ToolError,
+		ToolResult:   c.ToolResult,
+		CacheControl: fromLLMCache(c.Cache),
+	}
+}
+
+func fromLLMToolUse(tu *llm.ToolUse) *toolUse {
+	if tu == nil {
+		return nil
+	}
+	return &toolUse{
+		ID:   tu.ID,
+		Name: tu.Name,
+	}
+}
+
+func fromLLMMessage(msg llm.Message) message {
+	return message{
+		Role:    fromLLMRole[msg.Role],
+		Content: mapped(msg.Content, fromLLMContent),
+		ToolUse: fromLLMToolUse(msg.ToolUse),
+	}
+}
+
+func fromLLMToolChoice(tc *llm.ToolChoice) *toolChoice {
+	if tc == nil {
+		return nil
+	}
+	return &toolChoice{
+		Type: fromLLMToolChoiceType[tc.Type],
+		Name: tc.Name,
+	}
+}
+
+func fromLLMTool(t *llm.Tool) *tool {
+	return &tool{
+		Name:        t.Name,
+		Type:        t.Type,
+		Description: t.Description,
+		InputSchema: t.InputSchema,
+	}
+}
+
+func fromLLMSystem(s llm.SystemContent) systemContent {
+	return systemContent{
+		Text:         s.Text,
+		Type:         s.Type,
+		CacheControl: fromLLMCache(s.Cache),
+	}
+}
+
+func (s *Service) fromLLMRequest(r *llm.Request) *request {
+	return &request{
+		Model:      cmp.Or(s.Model, DefaultModel),
+		Messages:   mapped(r.Messages, fromLLMMessage),
+		MaxTokens:  cmp.Or(s.MaxTokens, DefaultMaxTokens),
+		ToolChoice: fromLLMToolChoice(r.ToolChoice),
+		Tools:      mapped(r.Tools, fromLLMTool),
+		System:     mapped(r.System, fromLLMSystem),
+	}
+}
+
+func toLLMUsage(u usage) llm.Usage {
+	return llm.Usage{
+		InputTokens:              u.InputTokens,
+		CacheCreationInputTokens: u.CacheCreationInputTokens,
+		CacheReadInputTokens:     u.CacheReadInputTokens,
+		OutputTokens:             u.OutputTokens,
+		CostUSD:                  u.CostUSD,
+	}
+}
+
+func toLLMContent(c content) llm.Content {
+	return llm.Content{
+		ID:         c.ID,
+		Type:       toLLMContentType[c.Type],
+		Text:       c.Text,
+		Thinking:   c.Thinking,
+		Data:       c.Data,
+		Signature:  c.Signature,
+		ToolName:   c.ToolName,
+		ToolInput:  c.ToolInput,
+		ToolUseID:  c.ToolUseID,
+		ToolError:  c.ToolError,
+		ToolResult: c.ToolResult,
+	}
+}
+
+func toLLMResponse(r *response) *llm.Response {
+	return &llm.Response{
+		ID:           r.ID,
+		Type:         r.Type,
+		Role:         toLLMRole[r.Role],
+		Model:        r.Model,
+		Content:      mapped(r.Content, toLLMContent),
+		StopReason:   toLLMStopReason[r.StopReason],
+		StopSequence: r.StopSequence,
+		Usage:        toLLMUsage(r.Usage),
+	}
+}
+
+// Do sends a request to Anthropic.
+func (s *Service) Do(ctx context.Context, ir *llm.Request) (*llm.Response, error) {
+	request := s.fromLLMRequest(ir)
+
+	var payload []byte
+	var err error
+	if dumpText || testing.Testing() {
+		payload, err = json.MarshalIndent(request, "", " ")
+	} else {
+		payload, err = json.Marshal(request)
+		payload = append(payload, '\n')
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	if false {
+		fmt.Printf("claude request payload:\n%s\n", payload)
+	}
+
+	backoff := []time.Duration{15 * time.Second, 30 * time.Second, time.Minute}
+	largerMaxTokens := false
+	var partialUsage usage
+
+	url := cmp.Or(s.URL, DefaultURL)
+	httpc := cmp.Or(s.HTTPC, http.DefaultClient)
+
+	// retry loop
+	for attempts := 0; ; attempts++ {
+		if dumpText {
+			fmt.Printf("RAW REQUEST:\n%s\n\n", payload)
+		}
+		req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(payload))
+		if err != nil {
+			return nil, err
+		}
+
+		req.Header.Set("Content-Type", "application/json")
+		req.Header.Set("X-API-Key", s.APIKey)
+		req.Header.Set("Anthropic-Version", "2023-06-01")
+
+		var features []string
+		if request.TokenEfficientToolUse {
+			features = append(features, "token-efficient-tool-use-2025-02-19")
+		}
+		if largerMaxTokens {
+			features = append(features, "output-128k-2025-02-19")
+			request.MaxTokens = 128 * 1024
+		}
+		if len(features) > 0 {
+			req.Header.Set("anthropic-beta", strings.Join(features, ","))
+		}
+
+		resp, err := httpc.Do(req)
+		if err != nil {
+			return nil, err
+		}
+		buf, _ := io.ReadAll(resp.Body)
+		resp.Body.Close()
+
+		switch {
+		case resp.StatusCode == http.StatusOK:
+			if dumpText {
+				fmt.Printf("RAW RESPONSE:\n%s\n\n", buf)
+			}
+			var response response
+			err = json.NewDecoder(bytes.NewReader(buf)).Decode(&response)
+			if err != nil {
+				return nil, err
+			}
+			if response.StopReason == "max_tokens" && !largerMaxTokens {
+				fmt.Printf("Retrying Anthropic API call with larger max tokens size.")
+				// Retry with more output tokens.
+				largerMaxTokens = true
+				response.Usage.CostUSD = response.TotalDollars()
+				partialUsage = response.Usage
+				continue
+			}
+
+			// Calculate and set the cost_usd field
+			if largerMaxTokens {
+				response.Usage.Add(partialUsage)
+			}
+			response.Usage.CostUSD = response.TotalDollars()
+
+			return toLLMResponse(&response), nil
+		case resp.StatusCode >= 500 && resp.StatusCode < 600:
+			// overloaded or unhappy, in one form or another
+			sleep := backoff[min(attempts, len(backoff)-1)] + time.Duration(rand.Int64N(int64(time.Second)))
+			slog.WarnContext(ctx, "anthropic_request_failed", "response", string(buf), "status_code", resp.StatusCode, "sleep", sleep)
+			time.Sleep(sleep)
+		case resp.StatusCode == 429:
+			// rate limited. wait 1 minute as a starting point, because that's the rate limiting window.
+			// and then add some additional time for backoff.
+			sleep := time.Minute + backoff[min(attempts, len(backoff)-1)] + time.Duration(rand.Int64N(int64(time.Second)))
+			slog.WarnContext(ctx, "anthropic_request_rate_limited", "response", string(buf), "sleep", sleep)
+			time.Sleep(sleep)
+		// case resp.StatusCode == 400:
+		// TODO: parse ErrorResponse, make (*ErrorResponse) implement error
+		default:
+			return nil, fmt.Errorf("API request failed with status %s\n%s", resp.Status, buf)
+		}
+	}
+}
+
+// cents per million tokens
+// (not dollars because i'm twitchy about using floats for money)
+type centsPer1MTokens struct {
+	Input         uint64
+	Output        uint64
+	CacheRead     uint64
+	CacheCreation uint64
+}
+
+// https://www.anthropic.com/pricing#anthropic-api
+var modelCost = map[string]centsPer1MTokens{
+	Claude37Sonnet: {
+		Input:         300,  // $3
+		Output:        1500, // $15
+		CacheRead:     30,   // $0.30
+		CacheCreation: 375,  // $3.75
+	},
+	Claude35Haiku: {
+		Input:         80,  // $0.80
+		Output:        400, // $4.00
+		CacheRead:     8,   // $0.08
+		CacheCreation: 100, // $1.00
+	},
+	Claude35Sonnet: {
+		Input:         300,  // $3
+		Output:        1500, // $15
+		CacheRead:     30,   // $0.30
+		CacheCreation: 375,  // $3.75
+	},
+}
+
+// TotalDollars returns the total cost to obtain this response, in dollars.
+func (mr *response) TotalDollars() float64 {
+	cpm, ok := modelCost[mr.Model]
+	if !ok {
+		panic(fmt.Sprintf("no pricing info for model: %s", mr.Model))
+	}
+	use := mr.Usage
+	megaCents := use.InputTokens*cpm.Input +
+		use.OutputTokens*cpm.Output +
+		use.CacheReadInputTokens*cpm.CacheRead +
+		use.CacheCreationInputTokens*cpm.CacheCreation
+	cents := float64(megaCents) / 1_000_000.0
+	return cents / 100.0
+}
diff --git a/llm/ant/ant_test.go b/llm/ant/ant_test.go
new file mode 100644
index 0000000..67cc5db
--- /dev/null
+++ b/llm/ant/ant_test.go
@@ -0,0 +1,93 @@
+package ant
+
+import (
+	"math"
+	"testing"
+)
+
+// TestCalculateCostFromTokens tests the calculateCostFromTokens function
+func TestCalculateCostFromTokens(t *testing.T) {
+	tests := []struct {
+		name                     string
+		model                    string
+		inputTokens              uint64
+		outputTokens             uint64
+		cacheReadInputTokens     uint64
+		cacheCreationInputTokens uint64
+		want                     float64
+	}{
+		{
+			name:                     "Zero tokens",
+			model:                    Claude37Sonnet,
+			inputTokens:              0,
+			outputTokens:             0,
+			cacheReadInputTokens:     0,
+			cacheCreationInputTokens: 0,
+			want:                     0,
+		},
+		{
+			name:                     "1000 input tokens, 500 output tokens",
+			model:                    Claude37Sonnet,
+			inputTokens:              1000,
+			outputTokens:             500,
+			cacheReadInputTokens:     0,
+			cacheCreationInputTokens: 0,
+			want:                     0.0105,
+		},
+		{
+			name:                     "10000 input tokens, 5000 output tokens",
+			model:                    Claude37Sonnet,
+			inputTokens:              10000,
+			outputTokens:             5000,
+			cacheReadInputTokens:     0,
+			cacheCreationInputTokens: 0,
+			want:                     0.105,
+		},
+		{
+			name:                     "With cache read tokens",
+			model:                    Claude37Sonnet,
+			inputTokens:              1000,
+			outputTokens:             500,
+			cacheReadInputTokens:     2000,
+			cacheCreationInputTokens: 0,
+			want:                     0.0111,
+		},
+		{
+			name:                     "With cache creation tokens",
+			model:                    Claude37Sonnet,
+			inputTokens:              1000,
+			outputTokens:             500,
+			cacheReadInputTokens:     0,
+			cacheCreationInputTokens: 1500,
+			want:                     0.016125,
+		},
+		{
+			name:                     "With all token types",
+			model:                    Claude37Sonnet,
+			inputTokens:              1000,
+			outputTokens:             500,
+			cacheReadInputTokens:     2000,
+			cacheCreationInputTokens: 1500,
+			want:                     0.016725,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			usage := usage{
+				InputTokens:              tt.inputTokens,
+				OutputTokens:             tt.outputTokens,
+				CacheReadInputTokens:     tt.cacheReadInputTokens,
+				CacheCreationInputTokens: tt.cacheCreationInputTokens,
+			}
+			mr := response{
+				Model: tt.model,
+				Usage: usage,
+			}
+			totalCost := mr.TotalDollars()
+			if math.Abs(totalCost-tt.want) > 0.0001 {
+				t.Errorf("totalCost = %v, want %v", totalCost, tt.want)
+			}
+		})
+	}
+}
diff --git a/llm/conversation/convo.go b/llm/conversation/convo.go
new file mode 100644
index 0000000..5a12256
--- /dev/null
+++ b/llm/conversation/convo.go
@@ -0,0 +1,617 @@
+package conversation
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"maps"
+	"math/rand/v2"
+	"slices"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/oklog/ulid/v2"
+	"github.com/richardlehane/crock32"
+	"sketch.dev/llm"
+	"sketch.dev/skribe"
+)
+
+type Listener interface {
+	// TODO: Content is leaking an anthropic API; should we avoid it?
+	// TODO: Where should we include start/end time and usage?
+	OnToolCall(ctx context.Context, convo *Convo, toolCallID string, toolName string, toolInput json.RawMessage, content llm.Content)
+	OnToolResult(ctx context.Context, convo *Convo, toolCallID string, toolName string, toolInput json.RawMessage, content llm.Content, result *string, err error)
+	OnRequest(ctx context.Context, convo *Convo, requestID string, msg *llm.Message)
+	OnResponse(ctx context.Context, convo *Convo, requestID string, msg *llm.Response)
+}
+
+type NoopListener struct{}
+
+func (n *NoopListener) OnToolCall(ctx context.Context, convo *Convo, id string, toolName string, toolInput json.RawMessage, content llm.Content) {
+}
+
+func (n *NoopListener) OnToolResult(ctx context.Context, convo *Convo, id string, toolName string, toolInput json.RawMessage, content llm.Content, result *string, err error) {
+}
+
+func (n *NoopListener) OnResponse(ctx context.Context, convo *Convo, id string, msg *llm.Response) {
+}
+func (n *NoopListener) OnRequest(ctx context.Context, convo *Convo, id string, msg *llm.Message) {}
+
+var ErrDoNotRespond = errors.New("do not respond")
+
+// A Convo is a managed conversation with Claude.
+// It automatically manages the state of the conversation,
+// including appending messages send/received,
+// calling tools and sending their results,
+// tracking usage, etc.
+//
+// Exported fields must not be altered concurrently with calling any method on Convo.
+// Typical usage is to configure a Convo once before using it.
+type Convo struct {
+	// ID is a unique ID for the conversation
+	ID string
+	// Ctx is the context for the entire conversation.
+	Ctx context.Context
+	// Service is the LLM service to use.
+	Service llm.Service
+	// Tools are the tools available during the conversation.
+	Tools []*llm.Tool
+	// SystemPrompt is the system prompt for the conversation.
+	SystemPrompt string
+	// PromptCaching indicates whether to use Anthropic's prompt caching.
+	// See https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching#continuing-a-multi-turn-conversation
+	// for the documentation. At request send time, we set the cache_control field on the
+	// last message. We also cache the system prompt.
+	// Default: true.
+	PromptCaching bool
+	// ToolUseOnly indicates whether Claude may only use tools during this conversation.
+	// TODO: add more fine-grained control over tool use?
+	ToolUseOnly bool
+	// Parent is the parent conversation, if any.
+	// It is non-nil for "subagent" calls.
+	// It is set automatically when calling SubConvo,
+	// and usually should not be set manually.
+	Parent *Convo
+	// Budget is the budget for this conversation (and all sub-conversations).
+	// The Conversation DOES NOT automatically enforce the budget.
+	// It is up to the caller to call OverBudget() as appropriate.
+	Budget Budget
+
+	// messages tracks the messages so far in the conversation.
+	messages []llm.Message
+
+	// Listener receives messages being sent.
+	Listener Listener
+
+	muToolUseCancel *sync.Mutex
+	toolUseCancel   map[string]context.CancelCauseFunc
+
+	// Protects usage. This is used for subconversations (that share part of CumulativeUsage) as well.
+	mu *sync.Mutex
+	// usage tracks usage for this conversation and all sub-conversations.
+	usage *CumulativeUsage
+}
+
+// newConvoID generates a new 8-byte random id.
+// The uniqueness/collision requirements here are very low.
+// They are not global identifiers,
+// just enough to distinguish different convos in a single session.
+func newConvoID() string {
+	u1 := rand.Uint32()
+	s := crock32.Encode(uint64(u1))
+	if len(s) < 7 {
+		s += strings.Repeat("0", 7-len(s))
+	}
+	return s[:3] + "-" + s[3:]
+}
+
+// New creates a new conversation with Claude with sensible defaults.
+// ctx is the context for the entire conversation.
+func New(ctx context.Context, srv llm.Service) *Convo {
+	id := newConvoID()
+	return &Convo{
+		Ctx:             skribe.ContextWithAttr(ctx, slog.String("convo_id", id)),
+		Service:         srv,
+		PromptCaching:   true,
+		usage:           newUsage(),
+		Listener:        &NoopListener{},
+		ID:              id,
+		muToolUseCancel: &sync.Mutex{},
+		toolUseCancel:   map[string]context.CancelCauseFunc{},
+		mu:              &sync.Mutex{},
+	}
+}
+
+// SubConvo creates a sub-conversation with the same configuration as the parent conversation.
+// (This propagates context for cancellation, HTTP client, API key, etc.)
+// The sub-conversation shares no messages with the parent conversation.
+// It does not inherit tools from the parent conversation.
+func (c *Convo) SubConvo() *Convo {
+	id := newConvoID()
+	return &Convo{
+		Ctx:           skribe.ContextWithAttr(c.Ctx, slog.String("convo_id", id), slog.String("parent_convo_id", c.ID)),
+		Service:       c.Service,
+		PromptCaching: c.PromptCaching,
+		Parent:        c,
+		// For convenience, sub-convo usage shares tool uses map with parent,
+		// all other fields separate, propagated in AddResponse
+		usage:    newUsageWithSharedToolUses(c.usage),
+		mu:       c.mu,
+		Listener: c.Listener,
+		ID:       id,
+		// Do not copy Budget. Each budget is independent,
+		// and OverBudget checks whether any ancestor is over budget.
+	}
+}
+
+func (c *Convo) SubConvoWithHistory() *Convo {
+	id := newConvoID()
+	return &Convo{
+		Ctx:           skribe.ContextWithAttr(c.Ctx, slog.String("convo_id", id), slog.String("parent_convo_id", c.ID)),
+		Service:       c.Service,
+		PromptCaching: c.PromptCaching,
+		Parent:        c,
+		// For convenience, sub-convo usage shares tool uses map with parent,
+		// all other fields separate, propagated in AddResponse
+		usage:    newUsageWithSharedToolUses(c.usage),
+		mu:       c.mu,
+		Listener: c.Listener,
+		ID:       id,
+		// Do not copy Budget. Each budget is independent,
+		// and OverBudget checks whether any ancestor is over budget.
+		messages: slices.Clone(c.messages),
+	}
+}
+
+// Depth reports how many "sub-conversations" deep this conversation is.
+// That it, it walks up parents until it finds a root.
+func (c *Convo) Depth() int {
+	x := c
+	var depth int
+	for x.Parent != nil {
+		x = x.Parent
+		depth++
+	}
+	return depth
+}
+
+// SendUserTextMessage sends a text message to the LLM in this conversation.
+// otherContents contains additional contents to send with the message, usually tool results.
+func (c *Convo) SendUserTextMessage(s string, otherContents ...llm.Content) (*llm.Response, error) {
+	contents := slices.Clone(otherContents)
+	if s != "" {
+		contents = append(contents, llm.Content{Type: llm.ContentTypeText, Text: s})
+	}
+	msg := llm.Message{
+		Role:    llm.MessageRoleUser,
+		Content: contents,
+	}
+	return c.SendMessage(msg)
+}
+
+func (c *Convo) messageRequest(msg llm.Message) *llm.Request {
+	system := []llm.SystemContent{}
+	if c.SystemPrompt != "" {
+		var d llm.SystemContent
+		d = llm.SystemContent{Type: "text", Text: c.SystemPrompt}
+		if c.PromptCaching {
+			d.Cache = true
+		}
+		system = []llm.SystemContent{d}
+	}
+
+	// Claude is happy to return an empty response in response to our Done() call,
+	// and, if so, you'll see something like:
+	// API request failed with status 400 Bad Request
+	// {"type":"error","error":  {"type":"invalid_request_error",
+	// "message":"messages.5: all messages must have non-empty content except for the optional final assistant message"}}
+	// So, we filter out those empty messages.
+	var nonEmptyMessages []llm.Message
+	for _, m := range c.messages {
+		if len(m.Content) > 0 {
+			nonEmptyMessages = append(nonEmptyMessages, m)
+		}
+	}
+
+	mr := &llm.Request{
+		Messages: append(nonEmptyMessages, msg), // not yet committed to keeping msg
+		System:   system,
+		Tools:    c.Tools,
+	}
+	if c.ToolUseOnly {
+		mr.ToolChoice = &llm.ToolChoice{Type: llm.ToolChoiceTypeAny}
+	}
+	return mr
+}
+
+func (c *Convo) findTool(name string) (*llm.Tool, error) {
+	for _, tool := range c.Tools {
+		if tool.Name == name {
+			return tool, nil
+		}
+	}
+	return nil, fmt.Errorf("tool %q not found", name)
+}
+
+// insertMissingToolResults adds error results for tool uses that were requested
+// but not included in the message, which can happen in error paths like "out of budget."
+// We only insert these if there were no tool responses at all, since an incorrect
+// number of tool results would be a programmer error. Mutates inputs.
+func (c *Convo) insertMissingToolResults(mr *llm.Request, msg *llm.Message) {
+	if len(mr.Messages) < 2 {
+		return
+	}
+	prev := mr.Messages[len(mr.Messages)-2]
+	var toolUsePrev int
+	for _, c := range prev.Content {
+		if c.Type == llm.ContentTypeToolUse {
+			toolUsePrev++
+		}
+	}
+	if toolUsePrev == 0 {
+		return
+	}
+	var toolUseCurrent int
+	for _, c := range msg.Content {
+		if c.Type == llm.ContentTypeToolResult {
+			toolUseCurrent++
+		}
+	}
+	if toolUseCurrent != 0 {
+		return
+	}
+	var prefix []llm.Content
+	for _, part := range prev.Content {
+		if part.Type != llm.ContentTypeToolUse {
+			continue
+		}
+		content := llm.Content{
+			Type:       llm.ContentTypeToolResult,
+			ToolUseID:  part.ID,
+			ToolError:  true,
+			ToolResult: "not executed; retry possible",
+		}
+		prefix = append(prefix, content)
+		msg.Content = append(prefix, msg.Content...)
+		mr.Messages[len(mr.Messages)-1].Content = msg.Content
+	}
+	slog.DebugContext(c.Ctx, "inserted missing tool results")
+}
+
+// SendMessage sends a message to Claude.
+// The conversation records (internally) all messages succesfully sent and received.
+func (c *Convo) SendMessage(msg llm.Message) (*llm.Response, error) {
+	id := ulid.Make().String()
+	mr := c.messageRequest(msg)
+	var lastMessage *llm.Message
+	if c.PromptCaching {
+		lastMessage = &mr.Messages[len(mr.Messages)-1]
+		if len(lastMessage.Content) > 0 {
+			lastMessage.Content[len(lastMessage.Content)-1].Cache = true
+		}
+	}
+	defer func() {
+		if lastMessage == nil {
+			return
+		}
+		if len(lastMessage.Content) > 0 {
+			lastMessage.Content[len(lastMessage.Content)-1].Cache = false
+		}
+	}()
+	c.insertMissingToolResults(mr, &msg)
+	c.Listener.OnRequest(c.Ctx, c, id, &msg)
+
+	startTime := time.Now()
+	resp, err := c.Service.Do(c.Ctx, mr)
+	if resp != nil {
+		resp.StartTime = &startTime
+		endTime := time.Now()
+		resp.EndTime = &endTime
+	}
+
+	if err != nil {
+		c.Listener.OnResponse(c.Ctx, c, id, nil)
+		return nil, err
+	}
+	c.messages = append(c.messages, msg, resp.ToMessage())
+	// Propagate usage to all ancestors (including us).
+	for x := c; x != nil; x = x.Parent {
+		x.usage.Add(resp.Usage)
+	}
+	c.Listener.OnResponse(c.Ctx, c, id, resp)
+	return resp, err
+}
+
+type toolCallInfoKeyType string
+
+var toolCallInfoKey toolCallInfoKeyType
+
+type ToolCallInfo struct {
+	ToolUseID string
+	Convo     *Convo
+}
+
+func ToolCallInfoFromContext(ctx context.Context) ToolCallInfo {
+	v := ctx.Value(toolCallInfoKey)
+	i, _ := v.(ToolCallInfo)
+	return i
+}
+
+func (c *Convo) ToolResultCancelContents(resp *llm.Response) ([]llm.Content, error) {
+	if resp.StopReason != llm.StopReasonToolUse {
+		return nil, nil
+	}
+	var toolResults []llm.Content
+
+	for _, part := range resp.Content {
+		if part.Type != llm.ContentTypeToolUse {
+			continue
+		}
+		c.incrementToolUse(part.ToolName)
+
+		content := llm.Content{
+			Type:      llm.ContentTypeToolResult,
+			ToolUseID: part.ID,
+		}
+
+		content.ToolError = true
+		content.ToolResult = "user canceled this too_use"
+		toolResults = append(toolResults, content)
+	}
+	return toolResults, nil
+}
+
+// GetID returns the conversation ID
+func (c *Convo) GetID() string {
+	return c.ID
+}
+
+func (c *Convo) CancelToolUse(toolUseID string, err error) error {
+	c.muToolUseCancel.Lock()
+	defer c.muToolUseCancel.Unlock()
+	cancel, ok := c.toolUseCancel[toolUseID]
+	if !ok {
+		return fmt.Errorf("cannot cancel %s: no cancel function registered for this tool_use_id. All I have is %+v", toolUseID, c.toolUseCancel)
+	}
+	delete(c.toolUseCancel, toolUseID)
+	cancel(err)
+	return nil
+}
+
+func (c *Convo) newToolUseContext(ctx context.Context, toolUseID string) (context.Context, context.CancelFunc) {
+	c.muToolUseCancel.Lock()
+	defer c.muToolUseCancel.Unlock()
+	ctx, cancel := context.WithCancelCause(ctx)
+	c.toolUseCancel[toolUseID] = cancel
+	return ctx, func() { c.CancelToolUse(toolUseID, nil) }
+}
+
+// ToolResultContents runs all tool uses requested by the response and returns their results.
+// Cancelling ctx will cancel any running tool calls.
+func (c *Convo) ToolResultContents(ctx context.Context, resp *llm.Response) ([]llm.Content, error) {
+	if resp.StopReason != llm.StopReasonToolUse {
+		return nil, nil
+	}
+	// Extract all tool calls from the response, call the tools, and gather the results.
+	var wg sync.WaitGroup
+	toolResultC := make(chan llm.Content, len(resp.Content))
+	for _, part := range resp.Content {
+		if part.Type != llm.ContentTypeToolUse {
+			continue
+		}
+		c.incrementToolUse(part.ToolName)
+		startTime := time.Now()
+
+		c.Listener.OnToolCall(ctx, c, part.ID, part.ToolName, part.ToolInput, llm.Content{
+			Type:             llm.ContentTypeToolUse,
+			ToolUseID:        part.ID,
+			ToolUseStartTime: &startTime,
+		})
+
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+
+			content := llm.Content{
+				Type:             llm.ContentTypeToolResult,
+				ToolUseID:        part.ID,
+				ToolUseStartTime: &startTime,
+			}
+			sendErr := func(err error) {
+				// Record end time
+				endTime := time.Now()
+				content.ToolUseEndTime = &endTime
+
+				content.ToolError = true
+				content.ToolResult = err.Error()
+				c.Listener.OnToolResult(ctx, c, part.ID, part.ToolName, part.ToolInput, content, nil, err)
+				toolResultC <- content
+			}
+			sendRes := func(res string) {
+				// Record end time
+				endTime := time.Now()
+				content.ToolUseEndTime = &endTime
+
+				content.ToolResult = res
+				c.Listener.OnToolResult(ctx, c, part.ID, part.ToolName, part.ToolInput, content, &res, nil)
+				toolResultC <- content
+			}
+
+			tool, err := c.findTool(part.ToolName)
+			if err != nil {
+				sendErr(err)
+				return
+			}
+			// Create a new context for just this tool_use call, and register its
+			// cancel function so that it can be canceled individually.
+			toolUseCtx, cancel := c.newToolUseContext(ctx, part.ID)
+			defer cancel()
+			// TODO: move this into newToolUseContext?
+			toolUseCtx = context.WithValue(toolUseCtx, toolCallInfoKey, ToolCallInfo{ToolUseID: part.ID, Convo: c})
+			toolResult, err := tool.Run(toolUseCtx, part.ToolInput)
+			if errors.Is(err, ErrDoNotRespond) {
+				return
+			}
+			if toolUseCtx.Err() != nil {
+				sendErr(context.Cause(toolUseCtx))
+				return
+			}
+
+			if err != nil {
+				sendErr(err)
+				return
+			}
+			sendRes(toolResult)
+		}()
+	}
+	wg.Wait()
+	close(toolResultC)
+	var toolResults []llm.Content
+	for toolResult := range toolResultC {
+		toolResults = append(toolResults, toolResult)
+	}
+	if ctx.Err() != nil {
+		return nil, ctx.Err()
+	}
+	return toolResults, nil
+}
+
+func (c *Convo) incrementToolUse(name string) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	c.usage.ToolUses[name]++
+}
+
+// CumulativeUsage represents cumulative usage across a Convo, including all sub-conversations.
+type CumulativeUsage struct {
+	StartTime                time.Time      `json:"start_time"`
+	Responses                uint64         `json:"messages"` // count of responses
+	InputTokens              uint64         `json:"input_tokens"`
+	OutputTokens             uint64         `json:"output_tokens"`
+	CacheReadInputTokens     uint64         `json:"cache_read_input_tokens"`
+	CacheCreationInputTokens uint64         `json:"cache_creation_input_tokens"`
+	TotalCostUSD             float64        `json:"total_cost_usd"`
+	ToolUses                 map[string]int `json:"tool_uses"` // tool name -> number of uses
+}
+
+func newUsage() *CumulativeUsage {
+	return &CumulativeUsage{ToolUses: make(map[string]int), StartTime: time.Now()}
+}
+
+func newUsageWithSharedToolUses(parent *CumulativeUsage) *CumulativeUsage {
+	return &CumulativeUsage{ToolUses: parent.ToolUses, StartTime: time.Now()}
+}
+
+func (u *CumulativeUsage) Clone() CumulativeUsage {
+	v := *u
+	v.ToolUses = maps.Clone(u.ToolUses)
+	return v
+}
+
+func (c *Convo) CumulativeUsage() CumulativeUsage {
+	if c == nil {
+		return CumulativeUsage{}
+	}
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.usage.Clone()
+}
+
+func (u *CumulativeUsage) WallTime() time.Duration {
+	return time.Since(u.StartTime)
+}
+
+func (u *CumulativeUsage) DollarsPerHour() float64 {
+	hours := u.WallTime().Hours()
+	// Prevent division by very small numbers that could cause issues
+	if hours < 1e-6 {
+		return 0
+	}
+	return u.TotalCostUSD / hours
+}
+
+func (u *CumulativeUsage) Add(usage llm.Usage) {
+	u.Responses++
+	u.InputTokens += usage.InputTokens
+	u.OutputTokens += usage.OutputTokens
+	u.CacheReadInputTokens += usage.CacheReadInputTokens
+	u.CacheCreationInputTokens += usage.CacheCreationInputTokens
+	u.TotalCostUSD += usage.CostUSD
+}
+
+// TotalInputTokens returns the grand total cumulative input tokens in u.
+func (u *CumulativeUsage) TotalInputTokens() uint64 {
+	return u.InputTokens + u.CacheReadInputTokens + u.CacheCreationInputTokens
+}
+
+// Attr returns the cumulative usage as a slog.Attr with key "usage".
+func (u CumulativeUsage) Attr() slog.Attr {
+	elapsed := time.Since(u.StartTime)
+	return slog.Group("usage",
+		slog.Duration("wall_time", elapsed),
+		slog.Uint64("responses", u.Responses),
+		slog.Uint64("input_tokens", u.InputTokens),
+		slog.Uint64("output_tokens", u.OutputTokens),
+		slog.Uint64("cache_read_input_tokens", u.CacheReadInputTokens),
+		slog.Uint64("cache_creation_input_tokens", u.CacheCreationInputTokens),
+		slog.Float64("total_cost_usd", u.TotalCostUSD),
+		slog.Float64("dollars_per_hour", u.TotalCostUSD/elapsed.Hours()),
+		slog.Any("tool_uses", maps.Clone(u.ToolUses)),
+	)
+}
+
+// A Budget represents the maximum amount of resources that may be spent on a conversation.
+// Note that the default (zero) budget is unlimited.
+type Budget struct {
+	MaxResponses uint64        // if > 0, max number of iterations (=responses)
+	MaxDollars   float64       // if > 0, max dollars that may be spent
+	MaxWallTime  time.Duration // if > 0, max wall time that may be spent
+}
+
+// OverBudget returns an error if the convo (or any of its parents) has exceeded its budget.
+// TODO: document parent vs sub budgets, multiple errors, etc, once we know the desired behavior.
+func (c *Convo) OverBudget() error {
+	for x := c; x != nil; x = x.Parent {
+		if err := x.overBudget(); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// ResetBudget sets the budget to the passed in budget and
+// adjusts it by what's been used so far.
+func (c *Convo) ResetBudget(budget Budget) {
+	c.Budget = budget
+	if c.Budget.MaxDollars > 0 {
+		c.Budget.MaxDollars += c.CumulativeUsage().TotalCostUSD
+	}
+	if c.Budget.MaxResponses > 0 {
+		c.Budget.MaxResponses += c.CumulativeUsage().Responses
+	}
+	if c.Budget.MaxWallTime > 0 {
+		c.Budget.MaxWallTime += c.usage.WallTime()
+	}
+}
+
+func (c *Convo) overBudget() error {
+	usage := c.CumulativeUsage()
+	// TODO: stop before we exceed the budget instead of after?
+	// Top priority is money, then time, then response count.
+	var err error
+	cont := "Continuing to chat will reset the budget."
+	if c.Budget.MaxDollars > 0 && usage.TotalCostUSD >= c.Budget.MaxDollars {
+		err = errors.Join(err, fmt.Errorf("$%.2f spent, budget is $%.2f. %s", usage.TotalCostUSD, c.Budget.MaxDollars, cont))
+	}
+	if c.Budget.MaxWallTime > 0 && usage.WallTime() >= c.Budget.MaxWallTime {
+		err = errors.Join(err, fmt.Errorf("%v elapsed, budget is %v. %s", usage.WallTime().Truncate(time.Second), c.Budget.MaxWallTime.Truncate(time.Second), cont))
+	}
+	if c.Budget.MaxResponses > 0 && usage.Responses >= c.Budget.MaxResponses {
+		err = errors.Join(err, fmt.Errorf("%d responses received, budget is %d. %s", usage.Responses, c.Budget.MaxResponses, cont))
+	}
+	return err
+}
diff --git a/llm/conversation/convo_test.go b/llm/conversation/convo_test.go
new file mode 100644
index 0000000..3fb1750
--- /dev/null
+++ b/llm/conversation/convo_test.go
@@ -0,0 +1,139 @@
+package conversation
+
+import (
+	"cmp"
+	"context"
+	"net/http"
+	"os"
+	"strings"
+	"testing"
+
+	"sketch.dev/httprr"
+	"sketch.dev/llm/ant"
+)
+
+func TestBasicConvo(t *testing.T) {
+	ctx := context.Background()
+	rr, err := httprr.Open("testdata/basic_convo.httprr", http.DefaultTransport)
+	if err != nil {
+		t.Fatal(err)
+	}
+	rr.ScrubReq(func(req *http.Request) error {
+		req.Header.Del("x-api-key")
+		return nil
+	})
+
+	apiKey := cmp.Or(os.Getenv("OUTER_SKETCH_ANTHROPIC_API_KEY"), os.Getenv("ANTHROPIC_API_KEY"))
+	srv := &ant.Service{
+		APIKey: apiKey,
+		HTTPC:  rr.Client(),
+	}
+	convo := New(ctx, srv)
+
+	const name = "Cornelius"
+	res, err := convo.SendUserTextMessage("Hi, my name is " + name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, part := range res.Content {
+		t.Logf("%s", part.Text)
+	}
+	res, err = convo.SendUserTextMessage("What is my name?")
+	if err != nil {
+		t.Fatal(err)
+	}
+	got := ""
+	for _, part := range res.Content {
+		got += part.Text
+	}
+	if !strings.Contains(got, name) {
+		t.Errorf("model does not know the given name %s: %q", name, got)
+	}
+}
+
+// TestCancelToolUse tests the CancelToolUse function of the Convo struct
+func TestCancelToolUse(t *testing.T) {
+	tests := []struct {
+		name         string
+		setupToolUse bool
+		toolUseID    string
+		cancelErr    error
+		expectError  bool
+		expectCancel bool
+	}{
+		{
+			name:         "Cancel existing tool use",
+			setupToolUse: true,
+			toolUseID:    "tool123",
+			cancelErr:    nil,
+			expectError:  false,
+			expectCancel: true,
+		},
+		{
+			name:         "Cancel existing tool use with error",
+			setupToolUse: true,
+			toolUseID:    "tool456",
+			cancelErr:    context.Canceled,
+			expectError:  false,
+			expectCancel: true,
+		},
+		{
+			name:         "Cancel non-existent tool use",
+			setupToolUse: false,
+			toolUseID:    "tool789",
+			cancelErr:    nil,
+			expectError:  true,
+			expectCancel: false,
+		},
+	}
+
+	srv := &ant.Service{}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			convo := New(context.Background(), srv)
+
+			var cancelCalled bool
+			var cancelledWithErr error
+
+			if tt.setupToolUse {
+				// Setup a mock cancel function to track calls
+				mockCancel := func(err error) {
+					cancelCalled = true
+					cancelledWithErr = err
+				}
+
+				convo.muToolUseCancel.Lock()
+				convo.toolUseCancel[tt.toolUseID] = mockCancel
+				convo.muToolUseCancel.Unlock()
+			}
+
+			err := convo.CancelToolUse(tt.toolUseID, tt.cancelErr)
+
+			// Check if we got the expected error state
+			if (err != nil) != tt.expectError {
+				t.Errorf("CancelToolUse() error = %v, expectError %v", err, tt.expectError)
+			}
+
+			// Check if the cancel function was called as expected
+			if cancelCalled != tt.expectCancel {
+				t.Errorf("Cancel function called = %v, expectCancel %v", cancelCalled, tt.expectCancel)
+			}
+
+			// If we expected the cancel to be called, verify it was called with the right error
+			if tt.expectCancel && cancelledWithErr != tt.cancelErr {
+				t.Errorf("Cancel function called with error = %v, expected %v", cancelledWithErr, tt.cancelErr)
+			}
+
+			// Verify the toolUseID was removed from the map if it was initially added
+			if tt.setupToolUse {
+				convo.muToolUseCancel.Lock()
+				_, exists := convo.toolUseCancel[tt.toolUseID]
+				convo.muToolUseCancel.Unlock()
+
+				if exists {
+					t.Errorf("toolUseID %s still exists in the map after cancellation", tt.toolUseID)
+				}
+			}
+		})
+	}
+}
diff --git a/llm/conversation/testdata/basic_convo.httprr b/llm/conversation/testdata/basic_convo.httprr
new file mode 100644
index 0000000..663de8d
--- /dev/null
+++ b/llm/conversation/testdata/basic_convo.httprr
@@ -0,0 +1,116 @@
+httprr trace v1
+457 1329
+POST https://api.anthropic.com/v1/messages HTTP/1.1
+Host: api.anthropic.com
+User-Agent: Go-http-client/1.1
+Content-Length: 261
+Anthropic-Version: 2023-06-01
+Content-Type: application/json
+
+{
+ "model": "claude-3-7-sonnet-20250219",
+ "messages": [
+  {
+   "role": "user",
+   "content": [
+    {
+     "type": "text",
+     "text": "Hi, my name is Cornelius",
+     "cache_control": {
+      "type": "ephemeral"
+     }
+    }
+   ]
+  }
+ ],
+ "max_tokens": 8192
+}HTTP/2.0 200 OK
+Anthropic-Organization-Id: 3c473a21-7208-450a-a9f8-80aebda45c1b
+Anthropic-Ratelimit-Input-Tokens-Limit: 200000
+Anthropic-Ratelimit-Input-Tokens-Remaining: 200000
+Anthropic-Ratelimit-Input-Tokens-Reset: 2025-03-11T17:45:06Z
+Anthropic-Ratelimit-Output-Tokens-Limit: 80000
+Anthropic-Ratelimit-Output-Tokens-Remaining: 79000
+Anthropic-Ratelimit-Output-Tokens-Reset: 2025-03-11T17:45:07Z
+Anthropic-Ratelimit-Requests-Limit: 4000
+Anthropic-Ratelimit-Requests-Remaining: 3999
+Anthropic-Ratelimit-Requests-Reset: 2025-03-11T17:45:05Z
+Anthropic-Ratelimit-Tokens-Limit: 280000
+Anthropic-Ratelimit-Tokens-Remaining: 279000
+Anthropic-Ratelimit-Tokens-Reset: 2025-03-11T17:45:06Z
+Cf-Cache-Status: DYNAMIC
+Cf-Ray: 91ecdd10fdc3f97f-SJC
+Content-Type: application/json
+Date: Tue, 11 Mar 2025 17:45:07 GMT
+Request-Id: req_01LBtxMdNzxDcDVPGJSh7giv
+Server: cloudflare
+Via: 1.1 google
+X-Robots-Tag: none
+
+{"id":"msg_01S1uUyUsTaKPBKuDUGGX8J2","type":"message","role":"assistant","model":"claude-3-7-sonnet-20250219","content":[{"type":"text","text":"Hello, Cornelius! It's nice to meet you. How are you doing today? Is there something I can help you with?"}],"stop_reason":"end_turn","stop_sequence":null,"usage":{"input_tokens":15,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":31}}779 1286
+POST https://api.anthropic.com/v1/messages HTTP/1.1
+Host: api.anthropic.com
+User-Agent: Go-http-client/1.1
+Content-Length: 583
+Anthropic-Version: 2023-06-01
+Content-Type: application/json
+
+{
+ "model": "claude-3-7-sonnet-20250219",
+ "messages": [
+  {
+   "role": "user",
+   "content": [
+    {
+     "type": "text",
+     "text": "Hi, my name is Cornelius"
+    }
+   ]
+  },
+  {
+   "role": "assistant",
+   "content": [
+    {
+     "type": "text",
+     "text": "Hello, Cornelius! It's nice to meet you. How are you doing today? Is there something I can help you with?"
+    }
+   ]
+  },
+  {
+   "role": "user",
+   "content": [
+    {
+     "type": "text",
+     "text": "What is my name?",
+     "cache_control": {
+      "type": "ephemeral"
+     }
+    }
+   ]
+  }
+ ],
+ "max_tokens": 8192
+}HTTP/2.0 200 OK
+Anthropic-Organization-Id: 3c473a21-7208-450a-a9f8-80aebda45c1b
+Anthropic-Ratelimit-Input-Tokens-Limit: 200000
+Anthropic-Ratelimit-Input-Tokens-Remaining: 200000
+Anthropic-Ratelimit-Input-Tokens-Reset: 2025-03-11T17:45:07Z
+Anthropic-Ratelimit-Output-Tokens-Limit: 80000
+Anthropic-Ratelimit-Output-Tokens-Remaining: 80000
+Anthropic-Ratelimit-Output-Tokens-Reset: 2025-03-11T17:45:07Z
+Anthropic-Ratelimit-Requests-Limit: 4000
+Anthropic-Ratelimit-Requests-Remaining: 3999
+Anthropic-Ratelimit-Requests-Reset: 2025-03-11T17:45:07Z
+Anthropic-Ratelimit-Tokens-Limit: 280000
+Anthropic-Ratelimit-Tokens-Remaining: 280000
+Anthropic-Ratelimit-Tokens-Reset: 2025-03-11T17:45:07Z
+Cf-Cache-Status: DYNAMIC
+Cf-Ray: 91ecdd1ae9a6f97f-SJC
+Content-Type: application/json
+Date: Tue, 11 Mar 2025 17:45:07 GMT
+Request-Id: req_01MBf3RWXNfQgwhVRwwkBYSn
+Server: cloudflare
+Via: 1.1 google
+X-Robots-Tag: none
+
+{"id":"msg_01FGz6DeWeDpspJG8cuxyVE9","type":"message","role":"assistant","model":"claude-3-7-sonnet-20250219","content":[{"type":"text","text":"Your name is Cornelius, as you mentioned in your introduction."}],"stop_reason":"end_turn","stop_sequence":null,"usage":{"input_tokens":54,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":17}}
\ No newline at end of file
diff --git a/llm/llm.go b/llm/llm.go
new file mode 100644
index 0000000..3ba6ed4
--- /dev/null
+++ b/llm/llm.go
@@ -0,0 +1,229 @@
+// Package llm provides a unified interface for interacting with LLMs.
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"log/slog"
+	"strings"
+	"time"
+)
+
+type Service interface {
+	// Do sends a request to an LLM.
+	Do(context.Context, *Request) (*Response, error)
+}
+
+// MustSchema validates that schema is a valid JSON schema and returns it as a json.RawMessage.
+// It panics if the schema is invalid.
+func MustSchema(schema string) json.RawMessage {
+	// TODO: validate schema, for now just make sure it's valid JSON
+	schema = strings.TrimSpace(schema)
+	bytes := []byte(schema)
+	if !json.Valid(bytes) {
+		panic("invalid JSON schema: " + schema)
+	}
+	return json.RawMessage(bytes)
+}
+
+type Request struct {
+	Messages   []Message
+	ToolChoice *ToolChoice
+	Tools      []*Tool
+	System     []SystemContent
+}
+
+// Message represents a message in the conversation.
+type Message struct {
+	Role    MessageRole
+	Content []Content
+	ToolUse *ToolUse // use to control whether/which tool to use
+}
+
+// ToolUse represents a tool use in the message content.
+type ToolUse struct {
+	ID   string
+	Name string
+}
+
+type ToolChoice struct {
+	Type ToolChoiceType
+	Name string
+}
+
+type SystemContent struct {
+	Text  string
+	Type  string
+	Cache bool
+}
+
+// Tool represents a tool available to an LLM.
+type Tool struct {
+	Name string
+	// Type is used by the text editor tool; see
+	// https://docs.anthropic.com/en/docs/build-with-claude/tool-use/text-editor-tool
+	Type        string
+	Description string
+	InputSchema json.RawMessage
+
+	// The Run function is automatically called when the tool is used.
+	// Run functions may be called concurrently with each other and themselves.
+	// The input to Run function is the input to the tool, as provided by Claude, in compliance with the input schema.
+	// The outputs from Run will be sent back to Claude.
+	// If you do not want to respond to the tool call request from Claude, return ErrDoNotRespond.
+	// ctx contains extra (rarely used) tool call information; retrieve it with ToolCallInfoFromContext.
+	Run func(ctx context.Context, input json.RawMessage) (string, error) `json:"-"`
+}
+
+type Content struct {
+	ID   string
+	Type ContentType
+	Text string
+
+	// for thinking
+	Thinking  string
+	Data      string
+	Signature string
+
+	// for tool_use
+	ToolName  string
+	ToolInput json.RawMessage
+
+	// for tool_result
+	ToolUseID  string
+	ToolError  bool
+	ToolResult string
+
+	// timing information for tool_result; added externally; not sent to the LLM
+	ToolUseStartTime *time.Time
+	ToolUseEndTime   *time.Time
+
+	Cache bool
+}
+
+func StringContent(s string) Content {
+	return Content{Type: ContentTypeText, Text: s}
+}
+
+// ContentsAttr returns contents as a slog.Attr.
+// It is meant for logging.
+func ContentsAttr(contents []Content) slog.Attr {
+	var contentAttrs []any // slog.Attr
+	for _, content := range contents {
+		var attrs []any // slog.Attr
+		switch content.Type {
+		case ContentTypeText:
+			attrs = append(attrs, slog.String("text", content.Text))
+		case ContentTypeToolUse:
+			attrs = append(attrs, slog.String("tool_name", content.ToolName))
+			attrs = append(attrs, slog.String("tool_input", string(content.ToolInput)))
+		case ContentTypeToolResult:
+			attrs = append(attrs, slog.String("tool_result", content.ToolResult))
+			attrs = append(attrs, slog.Bool("tool_error", content.ToolError))
+		case ContentTypeThinking:
+			attrs = append(attrs, slog.String("thinking", content.Text))
+		default:
+			attrs = append(attrs, slog.String("unknown_content_type", content.Type.String()))
+			attrs = append(attrs, slog.Any("text", content)) // just log it all raw, better to have too much than not enough
+		}
+		contentAttrs = append(contentAttrs, slog.Group(content.ID, attrs...))
+	}
+	return slog.Group("contents", contentAttrs...)
+}
+
+type (
+	MessageRole    int
+	ContentType    int
+	ToolChoiceType int
+	StopReason     int
+)
+
+//go:generate go tool golang.org/x/tools/cmd/stringer -type=MessageRole,ContentType,ToolChoiceType,StopReason -output=llm_string.go
+
+const (
+	MessageRoleUser MessageRole = iota
+	MessageRoleAssistant
+
+	ContentTypeText ContentType = iota
+	ContentTypeThinking
+	ContentTypeRedactedThinking
+	ContentTypeToolUse
+	ContentTypeToolResult
+
+	ToolChoiceTypeAuto ToolChoiceType = iota // default
+	ToolChoiceTypeAny                        // any tool, but must use one
+	ToolChoiceTypeNone                       // no tools allowed
+	ToolChoiceTypeTool                       // must use the tool specified in the Name field
+
+	StopReasonStopSequence StopReason = iota
+	StopReasonMaxTokens
+	StopReasonEndTurn
+	StopReasonToolUse
+)
+
+type Response struct {
+	ID           string
+	Type         string
+	Role         MessageRole
+	Model        string
+	Content      []Content
+	StopReason   StopReason
+	StopSequence *string
+	Usage        Usage
+	StartTime    *time.Time
+	EndTime      *time.Time
+}
+
+func (m *Response) ToMessage() Message {
+	return Message{
+		Role:    m.Role,
+		Content: m.Content,
+	}
+}
+
+// Usage represents the billing and rate-limit usage.
+// Most LLM structs do not have JSON tags, to avoid accidental direct use in specific providers.
+// However, the front-end uses this struct, and it relies on its JSON serialization.
+// Do NOT use this struct directly when implementing an llm.Service.
+type Usage struct {
+	InputTokens              uint64  `json:"input_tokens"`
+	CacheCreationInputTokens uint64  `json:"cache_creation_input_tokens"`
+	CacheReadInputTokens     uint64  `json:"cache_read_input_tokens"`
+	OutputTokens             uint64  `json:"output_tokens"`
+	CostUSD                  float64 `json:"cost_usd"`
+}
+
+func (u *Usage) Add(other Usage) {
+	u.InputTokens += other.InputTokens
+	u.CacheCreationInputTokens += other.CacheCreationInputTokens
+	u.CacheReadInputTokens += other.CacheReadInputTokens
+	u.OutputTokens += other.OutputTokens
+	u.CostUSD += other.CostUSD
+}
+
+func (u *Usage) String() string {
+	return fmt.Sprintf("in: %d, out: %d", u.InputTokens, u.OutputTokens)
+}
+
+func (u *Usage) IsZero() bool {
+	return *u == Usage{}
+}
+
+func (u *Usage) Attr() slog.Attr {
+	return slog.Group("usage",
+		slog.Uint64("input_tokens", u.InputTokens),
+		slog.Uint64("output_tokens", u.OutputTokens),
+		slog.Uint64("cache_creation_input_tokens", u.CacheCreationInputTokens),
+		slog.Uint64("cache_read_input_tokens", u.CacheReadInputTokens),
+		slog.Float64("cost_usd", u.CostUSD),
+	)
+}
+
+// UserStringMessage creates a user message with a single text content item.
+func UserStringMessage(text string) Message {
+	return Message{
+		Role:    MessageRoleUser,
+		Content: []Content{StringContent(text)},
+	}
+}
diff --git a/llm/llm_string.go b/llm/llm_string.go
new file mode 100644
index 0000000..1c3189e
--- /dev/null
+++ b/llm/llm_string.go
@@ -0,0 +1,88 @@
+// Code generated by "stringer -type=MessageRole,ContentType,ToolChoiceType,StopReason -output=llm_string.go"; DO NOT EDIT.
+
+package llm
+
+import "strconv"
+
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[MessageRoleUser-0]
+	_ = x[MessageRoleAssistant-1]
+}
+
+const _MessageRole_name = "MessageRoleUserMessageRoleAssistant"
+
+var _MessageRole_index = [...]uint8{0, 15, 35}
+
+func (i MessageRole) String() string {
+	if i < 0 || i >= MessageRole(len(_MessageRole_index)-1) {
+		return "MessageRole(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _MessageRole_name[_MessageRole_index[i]:_MessageRole_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[ContentTypeText-2]
+	_ = x[ContentTypeThinking-3]
+	_ = x[ContentTypeRedactedThinking-4]
+	_ = x[ContentTypeToolUse-5]
+	_ = x[ContentTypeToolResult-6]
+}
+
+const _ContentType_name = "ContentTypeTextContentTypeThinkingContentTypeRedactedThinkingContentTypeToolUseContentTypeToolResult"
+
+var _ContentType_index = [...]uint8{0, 15, 34, 61, 79, 100}
+
+func (i ContentType) String() string {
+	i -= 2
+	if i < 0 || i >= ContentType(len(_ContentType_index)-1) {
+		return "ContentType(" + strconv.FormatInt(int64(i+2), 10) + ")"
+	}
+	return _ContentType_name[_ContentType_index[i]:_ContentType_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[ToolChoiceTypeAuto-7]
+	_ = x[ToolChoiceTypeAny-8]
+	_ = x[ToolChoiceTypeNone-9]
+	_ = x[ToolChoiceTypeTool-10]
+}
+
+const _ToolChoiceType_name = "ToolChoiceTypeAutoToolChoiceTypeAnyToolChoiceTypeNoneToolChoiceTypeTool"
+
+var _ToolChoiceType_index = [...]uint8{0, 18, 35, 53, 71}
+
+func (i ToolChoiceType) String() string {
+	i -= 7
+	if i < 0 || i >= ToolChoiceType(len(_ToolChoiceType_index)-1) {
+		return "ToolChoiceType(" + strconv.FormatInt(int64(i+7), 10) + ")"
+	}
+	return _ToolChoiceType_name[_ToolChoiceType_index[i]:_ToolChoiceType_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[StopReasonStopSequence-11]
+	_ = x[StopReasonMaxTokens-12]
+	_ = x[StopReasonEndTurn-13]
+	_ = x[StopReasonToolUse-14]
+}
+
+const _StopReason_name = "StopReasonStopSequenceStopReasonMaxTokensStopReasonEndTurnStopReasonToolUse"
+
+var _StopReason_index = [...]uint8{0, 22, 41, 58, 75}
+
+func (i StopReason) String() string {
+	i -= 11
+	if i < 0 || i >= StopReason(len(_StopReason_index)-1) {
+		return "StopReason(" + strconv.FormatInt(int64(i+11), 10) + ")"
+	}
+	return _StopReason_name[_StopReason_index[i]:_StopReason_index[i+1]]
+}
diff --git a/llm/oai/oai.go b/llm/oai/oai.go
new file mode 100644
index 0000000..3e772ab
--- /dev/null
+++ b/llm/oai/oai.go
@@ -0,0 +1,592 @@
+package oai
+
+import (
+	"cmp"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"math/rand/v2"
+	"net/http"
+	"time"
+
+	"github.com/sashabaranov/go-openai"
+	"sketch.dev/llm"
+)
+
+const (
+	DefaultMaxTokens = 8192
+
+	OpenAIURL    = "https://api.openai.com/v1"
+	FireworksURL = "https://api.fireworks.ai/inference/v1"
+	LlamaCPPURL  = "http://localhost:8080/v1"
+	TogetherURL  = "https://api.together.xyz/v1"
+	GeminiURL    = "https://generativelanguage.googleapis.com/v1beta/openai/"
+
+	// Environment variable names for API keys
+	OpenAIAPIKeyEnv    = "OPENAI_API_KEY"
+	FireworksAPIKeyEnv = "FIREWORKS_API_KEY"
+	TogetherAPIKeyEnv  = "TOGETHER_API_KEY"
+	GeminiAPIKeyEnv    = "GEMINI_API_KEY"
+)
+
+type Model struct {
+	UserName  string // provided by the user to identify this model (e.g. "gpt4.1")
+	ModelName string // provided to the service provide to specify which model to use (e.g. "gpt-4.1-2025-04-14")
+	URL       string
+	Cost      ModelCost
+	APIKeyEnv string // environment variable name for the API key
+}
+
+type ModelCost struct {
+	Input       uint64 // in cents per million tokens
+	CachedInput uint64 // in cents per million tokens
+	Output      uint64 // in cents per million tokens
+}
+
+var (
+	DefaultModel = GPT41
+
+	GPT41 = Model{
+		UserName:  "gpt4.1",
+		ModelName: "gpt-4.1-2025-04-14",
+		URL:       OpenAIURL,
+		Cost:      ModelCost{Input: 200, CachedInput: 50, Output: 800},
+		APIKeyEnv: OpenAIAPIKeyEnv,
+	}
+
+	Gemini25Flash = Model{
+		UserName:  "gemini-flash-2.5",
+		ModelName: "gemini-2.5-flash-preview-04-17",
+		URL:       GeminiURL,
+		Cost:      ModelCost{Input: 15, Output: 60},
+		APIKeyEnv: GeminiAPIKeyEnv,
+	}
+
+	Gemini25Pro = Model{
+		UserName:  "gemini-pro-2.5",
+		ModelName: "gemini-2.5-pro-preview-03-25",
+		URL:       GeminiURL,
+		// GRRRR. Really??
+		// Input is: $1.25, prompts <= 200k tokens, $2.50, prompts > 200k tokens
+		// Output is: $10.00, prompts <= 200k tokens, $15.00, prompts > 200k
+		// Caching is: $0.31, prompts <= 200k tokens, $0.625, prompts > 200k, $4.50 / 1,000,000 tokens per hour
+		// Whatever that means. Are we caching? I have no idea.
+		// How do you always manage to be the annoying one, Google?
+		// I'm not complicating things just for you.
+		Cost:      ModelCost{Input: 125, Output: 1000},
+		APIKeyEnv: GeminiAPIKeyEnv,
+	}
+
+	TogetherDeepseekV3 = Model{
+		UserName:  "together-deepseek-v3",
+		ModelName: "deepseek-ai/DeepSeek-V3",
+		URL:       TogetherURL,
+		Cost:      ModelCost{Input: 125, Output: 125},
+		APIKeyEnv: TogetherAPIKeyEnv,
+	}
+
+	TogetherLlama4Maverick = Model{
+		UserName:  "together-llama4-maverick",
+		ModelName: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+		URL:       TogetherURL,
+		Cost:      ModelCost{Input: 27, Output: 85},
+		APIKeyEnv: TogetherAPIKeyEnv,
+	}
+
+	TogetherLlama3_3_70B = Model{
+		UserName:  "together-llama3-70b",
+		ModelName: "meta-llama/Llama-3.3-70B-Instruct-Turbo",
+		URL:       TogetherURL,
+		Cost:      ModelCost{Input: 88, Output: 88},
+		APIKeyEnv: TogetherAPIKeyEnv,
+	}
+
+	TogetherMistralSmall = Model{
+		UserName:  "together-mistral-small",
+		ModelName: "mistralai/Mistral-Small-24B-Instruct-2501",
+		URL:       TogetherURL,
+		Cost:      ModelCost{Input: 80, Output: 80},
+		APIKeyEnv: TogetherAPIKeyEnv,
+	}
+
+	LlamaCPP = Model{
+		UserName:  "llama.cpp",
+		ModelName: "llama.cpp local model",
+		URL:       LlamaCPPURL,
+		// zero cost
+		Cost: ModelCost{},
+	}
+
+	FireworksDeepseekV3 = Model{
+		UserName:  "fireworks-deepseek-v3",
+		ModelName: "accounts/fireworks/models/deepseek-v3-0324",
+		URL:       FireworksURL,
+		Cost:      ModelCost{Input: 90, Output: 90}, // not entirely sure about this, they don't list pricing anywhere convenient
+		APIKeyEnv: FireworksAPIKeyEnv,
+	}
+)
+
+// Service provides chat completions.
+// Fields should not be altered concurrently with calling any method on Service.
+type Service struct {
+	HTTPC     *http.Client // defaults to http.DefaultClient if nil
+	APIKey    string       // optional, if not set will try to load from env var
+	Model     Model        // defaults to DefaultModel if zero value
+	MaxTokens int          // defaults to DefaultMaxTokens if zero
+	Org       string       // optional - organization ID
+}
+
+var _ llm.Service = (*Service)(nil)
+
+// ModelsRegistry is a registry of all known models with their user-friendly names.
+var ModelsRegistry = []Model{
+	GPT41,
+	Gemini25Flash,
+	Gemini25Pro,
+	TogetherDeepseekV3,
+	TogetherLlama4Maverick,
+	TogetherLlama3_3_70B,
+	TogetherMistralSmall,
+	LlamaCPP,
+	FireworksDeepseekV3,
+}
+
+// ListModels returns a list of all available models with their user-friendly names.
+func ListModels() []string {
+	var names []string
+	for _, model := range ModelsRegistry {
+		if model.UserName != "" {
+			names = append(names, model.UserName)
+		}
+	}
+	return names
+}
+
+// ModelByUserName returns a model by its user-friendly name.
+// Returns nil if no model with the given name is found.
+func ModelByUserName(name string) *Model {
+	for _, model := range ModelsRegistry {
+		if model.UserName == name {
+			return &model
+		}
+	}
+	return nil
+}
+
+var (
+	fromLLMRole = map[llm.MessageRole]string{
+		llm.MessageRoleAssistant: "assistant",
+		llm.MessageRoleUser:      "user",
+	}
+	fromLLMContentType = map[llm.ContentType]string{
+		llm.ContentTypeText:             "text",
+		llm.ContentTypeToolUse:          "function", // OpenAI uses function instead of tool_call
+		llm.ContentTypeToolResult:       "tool_result",
+		llm.ContentTypeThinking:         "text", // Map thinking to text since OpenAI doesn't have thinking
+		llm.ContentTypeRedactedThinking: "text", // Map redacted_thinking to text
+	}
+	fromLLMToolChoiceType = map[llm.ToolChoiceType]string{
+		llm.ToolChoiceTypeAuto: "auto",
+		llm.ToolChoiceTypeAny:  "any",
+		llm.ToolChoiceTypeNone: "none",
+		llm.ToolChoiceTypeTool: "function", // OpenAI uses "function" instead of "tool"
+	}
+	toLLMRole = map[string]llm.MessageRole{
+		"assistant": llm.MessageRoleAssistant,
+		"user":      llm.MessageRoleUser,
+	}
+	toLLMStopReason = map[string]llm.StopReason{
+		"stop":           llm.StopReasonStopSequence,
+		"length":         llm.StopReasonMaxTokens,
+		"tool_calls":     llm.StopReasonToolUse,
+		"function_call":  llm.StopReasonToolUse,      // Map both to ToolUse
+		"content_filter": llm.StopReasonStopSequence, // No direct equivalent
+	}
+)
+
+// fromLLMContent converts llm.Content to the format expected by OpenAI.
+func fromLLMContent(c llm.Content) (string, []openai.ToolCall) {
+	switch c.Type {
+	case llm.ContentTypeText:
+		return c.Text, nil
+	case llm.ContentTypeToolUse:
+		// For OpenAI, tool use is sent as a null content with tool_calls in the message
+		return "", []openai.ToolCall{
+			{
+				Type: openai.ToolTypeFunction,
+				ID:   c.ID, // Use the content ID if provided
+				Function: openai.FunctionCall{
+					Name:      c.ToolName,
+					Arguments: string(c.ToolInput),
+				},
+			},
+		}
+	case llm.ContentTypeToolResult:
+		// Tool results in OpenAI are sent as a separate message with tool_call_id
+		return c.ToolResult, nil
+	default:
+		// For thinking or other types, convert to text
+		return c.Text, nil
+	}
+}
+
+// fromLLMMessage converts llm.Message to OpenAI ChatCompletionMessage format
+func fromLLMMessage(msg llm.Message) []openai.ChatCompletionMessage {
+	// For OpenAI, we need to handle tool results differently than regular messages
+	// Each tool result becomes its own message with role="tool"
+
+	var messages []openai.ChatCompletionMessage
+
+	// Check if this is a regular message or contains tool results
+	var regularContent []llm.Content
+	var toolResults []llm.Content
+
+	for _, c := range msg.Content {
+		if c.Type == llm.ContentTypeToolResult {
+			toolResults = append(toolResults, c)
+		} else {
+			regularContent = append(regularContent, c)
+		}
+	}
+
+	// Process tool results as separate messages, but first
+	for _, tr := range toolResults {
+		m := openai.ChatCompletionMessage{
+			Role:       "tool",
+			Content:    cmp.Or(tr.ToolResult, " "), // TODO: remove omitempty upstream
+			ToolCallID: tr.ToolUseID,
+		}
+		messages = append(messages, m)
+	}
+	// Process regular content second
+	if len(regularContent) > 0 {
+		m := openai.ChatCompletionMessage{
+			Role: fromLLMRole[msg.Role],
+		}
+
+		// For assistant messages that contain tool calls
+		var toolCalls []openai.ToolCall
+		var textContent string
+
+		for _, c := range regularContent {
+			content, tools := fromLLMContent(c)
+			if len(tools) > 0 {
+				toolCalls = append(toolCalls, tools...)
+			} else if content != "" {
+				if textContent != "" {
+					textContent += "\n"
+				}
+				textContent += content
+			}
+		}
+
+		m.Content = textContent
+		m.ToolCalls = toolCalls
+
+		messages = append(messages, m)
+	}
+
+	return messages
+}
+
+// fromLLMToolChoice converts llm.ToolChoice to the format expected by OpenAI.
+func fromLLMToolChoice(tc *llm.ToolChoice) any {
+	if tc == nil {
+		return nil
+	}
+
+	if tc.Type == llm.ToolChoiceTypeTool && tc.Name != "" {
+		return openai.ToolChoice{
+			Type: openai.ToolTypeFunction,
+			Function: openai.ToolFunction{
+				Name: tc.Name,
+			},
+		}
+	}
+
+	// For non-specific tool choice, just use the string
+	return fromLLMToolChoiceType[tc.Type]
+}
+
+// fromLLMTool converts llm.Tool to the format expected by OpenAI.
+func fromLLMTool(t *llm.Tool) openai.Tool {
+	return openai.Tool{
+		Type: openai.ToolTypeFunction,
+		Function: &openai.FunctionDefinition{
+			Name:        t.Name,
+			Description: t.Description,
+			Parameters:  t.InputSchema,
+		},
+	}
+}
+
+// fromLLMSystem converts llm.SystemContent to an OpenAI system message.
+func fromLLMSystem(systemContent []llm.SystemContent) []openai.ChatCompletionMessage {
+	if len(systemContent) == 0 {
+		return nil
+	}
+
+	// Combine all system content into a single system message
+	var systemText string
+	for i, content := range systemContent {
+		if i > 0 && systemText != "" && content.Text != "" {
+			systemText += "\n"
+		}
+		systemText += content.Text
+	}
+
+	if systemText == "" {
+		return nil
+	}
+
+	return []openai.ChatCompletionMessage{
+		{
+			Role:    "system",
+			Content: systemText,
+		},
+	}
+}
+
+// toRawLLMContent converts a raw content string from OpenAI to llm.Content.
+func toRawLLMContent(content string) llm.Content {
+	return llm.Content{
+		Type: llm.ContentTypeText,
+		Text: content,
+	}
+}
+
+// toToolCallLLMContent converts a tool call from OpenAI to llm.Content.
+func toToolCallLLMContent(toolCall openai.ToolCall) llm.Content {
+	// Generate a content ID if needed
+	id := toolCall.ID
+	if id == "" {
+		// Create a deterministic ID based on the function name if no ID is provided
+		id = "tc_" + toolCall.Function.Name
+	}
+
+	return llm.Content{
+		ID:        id,
+		Type:      llm.ContentTypeToolUse,
+		ToolName:  toolCall.Function.Name,
+		ToolInput: json.RawMessage(toolCall.Function.Arguments),
+	}
+}
+
+// toToolResultLLMContent converts a tool result message from OpenAI to llm.Content.
+func toToolResultLLMContent(msg openai.ChatCompletionMessage) llm.Content {
+	return llm.Content{
+		Type:       llm.ContentTypeToolResult,
+		ToolUseID:  msg.ToolCallID,
+		ToolResult: msg.Content,
+		ToolError:  false, // OpenAI doesn't specify errors explicitly
+	}
+}
+
+// toLLMContents converts message content from OpenAI to []llm.Content.
+func toLLMContents(msg openai.ChatCompletionMessage) []llm.Content {
+	var contents []llm.Content
+
+	// If this is a tool response, handle it separately
+	if msg.Role == "tool" && msg.ToolCallID != "" {
+		return []llm.Content{toToolResultLLMContent(msg)}
+	}
+
+	// If there's text content, add it
+	if msg.Content != "" {
+		contents = append(contents, toRawLLMContent(msg.Content))
+	}
+
+	// If there are tool calls, add them
+	for _, tc := range msg.ToolCalls {
+		contents = append(contents, toToolCallLLMContent(tc))
+	}
+
+	// If empty, add an empty text content
+	if len(contents) == 0 {
+		contents = append(contents, llm.Content{
+			Type: llm.ContentTypeText,
+			Text: "",
+		})
+	}
+
+	return contents
+}
+
+// toLLMUsage converts usage information from OpenAI to llm.Usage.
+func (s *Service) toLLMUsage(model string, au openai.Usage) llm.Usage {
+	// fmt.Printf("raw usage: %+v / %v / %v\n", au, au.PromptTokensDetails, au.CompletionTokensDetails)
+	in := uint64(au.PromptTokens)
+	var inc uint64
+	if au.PromptTokensDetails != nil {
+		inc = uint64(au.PromptTokensDetails.CachedTokens)
+	}
+	out := uint64(au.CompletionTokens)
+	u := llm.Usage{
+		InputTokens:              in,
+		CacheReadInputTokens:     inc,
+		CacheCreationInputTokens: in,
+		OutputTokens:             out,
+	}
+	u.CostUSD = s.calculateCostFromTokens(u)
+	return u
+}
+
+// toLLMResponse converts the OpenAI response to llm.Response.
+func (s *Service) toLLMResponse(r *openai.ChatCompletionResponse) *llm.Response {
+	// fmt.Printf("Raw response\n")
+	// enc := json.NewEncoder(os.Stdout)
+	// enc.SetIndent("", "  ")
+	// enc.Encode(r)
+	// fmt.Printf("\n")
+
+	if len(r.Choices) == 0 {
+		return &llm.Response{
+			ID:    r.ID,
+			Model: r.Model,
+			Role:  llm.MessageRoleAssistant,
+			Usage: s.toLLMUsage(r.Model, r.Usage),
+		}
+	}
+
+	// Process the primary choice
+	choice := r.Choices[0]
+
+	return &llm.Response{
+		ID:         r.ID,
+		Model:      r.Model,
+		Role:       toRoleFromString(choice.Message.Role),
+		Content:    toLLMContents(choice.Message),
+		StopReason: toStopReason(string(choice.FinishReason)),
+		Usage:      s.toLLMUsage(r.Model, r.Usage),
+	}
+}
+
+// toRoleFromString converts a role string to llm.MessageRole.
+func toRoleFromString(role string) llm.MessageRole {
+	if role == "tool" || role == "system" || role == "function" {
+		return llm.MessageRoleAssistant // Map special roles to assistant for consistency
+	}
+	if mr, ok := toLLMRole[role]; ok {
+		return mr
+	}
+	return llm.MessageRoleUser // Default to user if unknown
+}
+
+// toStopReason converts a finish reason string to llm.StopReason.
+func toStopReason(reason string) llm.StopReason {
+	if sr, ok := toLLMStopReason[reason]; ok {
+		return sr
+	}
+	return llm.StopReasonStopSequence // Default
+}
+
+// calculateCostFromTokens calculates the cost in dollars for the given model and token counts.
+func (s *Service) calculateCostFromTokens(u llm.Usage) float64 {
+	cost := s.Model.Cost
+
+	// TODO: check this for correctness, i am skeptical
+	// Calculate cost in cents
+	megaCents := u.CacheCreationInputTokens*cost.Input +
+		u.CacheReadInputTokens*cost.CachedInput +
+		u.OutputTokens*cost.Output
+
+	cents := float64(megaCents) / 1_000_000
+	// Convert to dollars
+	dollars := cents / 100.0
+	// fmt.Printf("in_new=%d, in_cached=%d, out=%d, cost=%.2f\n", u.CacheCreationInputTokens, u.CacheReadInputTokens, u.OutputTokens, dollars)
+	return dollars
+}
+
+// Do sends a request to OpenAI using the go-openai package.
+func (s *Service) Do(ctx context.Context, ir *llm.Request) (*llm.Response, error) {
+	// Configure the OpenAI client
+	httpc := cmp.Or(s.HTTPC, http.DefaultClient)
+	model := cmp.Or(s.Model, DefaultModel)
+
+	// TODO: do this one during Service setup? maybe with a constructor instead?
+	config := openai.DefaultConfig(s.APIKey)
+	if model.URL != "" {
+		config.BaseURL = model.URL
+	}
+	if s.Org != "" {
+		config.OrgID = s.Org
+	}
+	config.HTTPClient = httpc
+
+	client := openai.NewClientWithConfig(config)
+
+	// Start with system messages if provided
+	var allMessages []openai.ChatCompletionMessage
+	if len(ir.System) > 0 {
+		sysMessages := fromLLMSystem(ir.System)
+		allMessages = append(allMessages, sysMessages...)
+	}
+
+	// Add regular and tool messages
+	for _, msg := range ir.Messages {
+		msgs := fromLLMMessage(msg)
+		allMessages = append(allMessages, msgs...)
+	}
+
+	// Convert tools
+	var tools []openai.Tool
+	for _, t := range ir.Tools {
+		tools = append(tools, fromLLMTool(t))
+	}
+
+	// Create the OpenAI request
+	req := openai.ChatCompletionRequest{
+		Model:      model.ModelName,
+		Messages:   allMessages,
+		MaxTokens:  cmp.Or(s.MaxTokens, DefaultMaxTokens),
+		Tools:      tools,
+		ToolChoice: fromLLMToolChoice(ir.ToolChoice), // TODO: make fromLLMToolChoice return an error when a perfect translation is not possible
+	}
+	// fmt.Printf("Sending request to OpenAI\n")
+	// enc := json.NewEncoder(os.Stdout)
+	// enc.SetIndent("", "  ")
+	// enc.Encode(req)
+	// fmt.Printf("\n")
+
+	// Retry mechanism
+	backoff := []time.Duration{1 * time.Second, 2 * time.Second, 5 * time.Second}
+
+	// retry loop
+	for attempts := 0; ; attempts++ {
+		resp, err := client.CreateChatCompletion(ctx, req)
+
+		// Handle successful response
+		if err == nil {
+			return s.toLLMResponse(&resp), nil
+		}
+
+		// Handle errors
+		var apiErr *openai.APIError
+		if ok := errors.As(err, &apiErr); !ok {
+			// Not an OpenAI API error, return immediately
+			return nil, err
+		}
+
+		switch {
+		case apiErr.HTTPStatusCode >= 500:
+			// Server error, try again with backoff
+			sleep := backoff[min(attempts, len(backoff)-1)] + time.Duration(rand.Int64N(int64(time.Second)))
+			slog.WarnContext(ctx, "openai_request_failed", "error", apiErr.Error(), "status_code", apiErr.HTTPStatusCode, "sleep", sleep)
+			time.Sleep(sleep)
+			continue
+
+		case apiErr.HTTPStatusCode == 429:
+			// Rate limited, back off longer
+			sleep := 20*time.Second + backoff[min(attempts, len(backoff)-1)] + time.Duration(rand.Int64N(int64(time.Second)))
+			slog.WarnContext(ctx, "openai_request_rate_limited", "error", apiErr.Error(), "sleep", sleep)
+			time.Sleep(sleep)
+			continue
+
+		default:
+			// Other error, return immediately
+			return nil, fmt.Errorf("OpenAI API error: %w", err)
+		}
+	}
+}
diff --git a/llm/oai/oai_test.go b/llm/oai/oai_test.go
new file mode 100644
index 0000000..7bea552
--- /dev/null
+++ b/llm/oai/oai_test.go
@@ -0,0 +1,96 @@
+package oai
+
+import (
+	"math"
+	"testing"
+
+	"sketch.dev/llm"
+)
+
+// TestCalculateCostFromTokens tests the calculateCostFromTokens method
+func TestCalculateCostFromTokens(t *testing.T) {
+	tests := []struct {
+		name                string
+		model               Model
+		cacheCreationTokens uint64
+		cacheReadTokens     uint64
+		outputTokens        uint64
+		want                float64
+	}{
+		{
+			name:                "Zero tokens",
+			model:               GPT41,
+			cacheCreationTokens: 0,
+			cacheReadTokens:     0,
+			outputTokens:        0,
+			want:                0,
+		},
+		{
+			name:                "1000 input tokens, 500 output tokens",
+			model:               GPT41,
+			cacheCreationTokens: 1000,
+			cacheReadTokens:     0,
+			outputTokens:        500,
+			// GPT41: Input: 200 per million, Output: 800 per million
+			// (1000 * 200 + 500 * 800) / 1_000_000 / 100 = 0.006
+			want: 0.006,
+		},
+		{
+			name:                "10000 input tokens, 5000 output tokens",
+			model:               GPT41,
+			cacheCreationTokens: 10000,
+			cacheReadTokens:     0,
+			outputTokens:        5000,
+			// (10000 * 200 + 5000 * 800) / 1_000_000 / 100 = 0.06
+			want: 0.06,
+		},
+		{
+			name:                "1000 input tokens, 500 output tokens Gemini",
+			model:               Gemini25Flash,
+			cacheCreationTokens: 1000,
+			cacheReadTokens:     0,
+			outputTokens:        500,
+			// Gemini25Flash: Input: 15 per million, Output: 60 per million
+			// (1000 * 15 + 500 * 60) / 1_000_000 / 100 = 0.00045
+			want: 0.00045,
+		},
+		{
+			name:                "With cache read tokens",
+			model:               GPT41,
+			cacheCreationTokens: 500,
+			cacheReadTokens:     500, // 500 tokens from cache
+			outputTokens:        500,
+			// (500 * 200 + 500 * 50 + 500 * 800) / 1_000_000 / 100 = 0.00525
+			want: 0.00525,
+		},
+		{
+			name:                "With all token types",
+			model:               GPT41,
+			cacheCreationTokens: 1000,
+			cacheReadTokens:     1000,
+			outputTokens:        1000,
+			// (1000 * 200 + 1000 * 50 + 1000 * 800) / 1_000_000 / 100 = 0.0105
+			want: 0.0105,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			// Create a service with the test model
+			svc := &Service{Model: tt.model}
+
+			// Create a usage object
+			usage := llm.Usage{
+				CacheCreationInputTokens: tt.cacheCreationTokens,
+				CacheReadInputTokens:     tt.cacheReadTokens,
+				OutputTokens:             tt.outputTokens,
+			}
+
+			totalCost := svc.calculateCostFromTokens(usage)
+			if math.Abs(totalCost-tt.want) > 0.0001 {
+				t.Errorf("calculateCostFromTokens(%s, cache_creation=%d, cache_read=%d, output=%d) = %v, want %v",
+					tt.model.ModelName, tt.cacheCreationTokens, tt.cacheReadTokens, tt.outputTokens, totalCost, tt.want)
+			}
+		})
+	}
+}