# llms-full.txt
# @amplitude/ai 0.3.3 — Complete API Reference for AI Coding Agents
#
# This file is the definitive guide for any coding agent instrumenting
# a JavaScript/TypeScript AI application with @amplitude/ai.
# For tool discovery only, see llms.txt.

## Quick Setup for AI Coding Agents

Install the SDK, then configure your coding agent's MCP server for
project analysis tools (scan_project, validate_file):

```bash
pnpm add @amplitude/ai
```

MCP config (one line — works with any MCP-compatible agent):
```json
{ "amplitude-ai": { "command": "npx", "args": ["amplitude-ai", "mcp"] } }
```

Agent-specific setup:

Cursor: Add to .cursor/mcp.json:
  { "mcpServers": { "amplitude-ai": { "command": "npx", "args": ["amplitude-ai", "mcp"] } } }
  Then point the agent at: node_modules/@amplitude/ai/amplitude-ai.md

Claude Code:
  claude mcp add amplitude-ai -- npx amplitude-ai mcp
  Then point the agent at: node_modules/@amplitude/ai/amplitude-ai.md

Codex CLI: Add to ~/.codex/config.toml:
  [mcp_servers.amplitude-ai]
  command = "npx"
  args = ["amplitude-ai", "mcp"]
  AGENTS.md in the npm package is auto-read by Codex.

Without MCP: Read amplitude-ai.md for the complete guided workflow, or this file for the API reference.

## Instrumentation Workflow (4 Phases)

Phase 1 — Detect: Run scan_project MCP tool (or read package.json + grep for providers manually).
Phase 2 — Discover: Identify agents, call sites, delegation patterns. Read source files.
Phase 3 — Instrument: Create bootstrap file, swap providers, add session wrapping, tool tracking.
Phase 4 — Verify: Run tsc --noEmit, validate_file on changed files, amplitude-ai doctor.

For the full guided workflow with code examples for every phase,
read the MCP resource: amplitude-ai://instrument-guide

---

## Core API

### AmplitudeAI(options)
Initialize the SDK. Required entry point.
```
import { AmplitudeAI, AIConfig } from '@amplitude/ai';
const ai = new AmplitudeAI({
  apiKey: string,
  config?: new AIConfig({
    contentMode?: 'full' | 'metadata_only' | 'customer_enriched',
    redactPii?: boolean,
    debug?: boolean,
  }),
});
```

### patch(options) / unpatch()
Zero-code instrumentation. Monkey-patches all detected provider SDKs.
```
import { patch, unpatch } from '@amplitude/ai';
patch({ amplitudeAI: ai });
// All provider calls are now tracked. Call unpatch() to restore.
```

### wrap(client, ai, opts?)
Wrap an existing provider client without changing its construction.
```
import { wrap } from '@amplitude/ai';
const instrumented = wrap(existingOpenAIClient, ai);
```

### ai.agent(agentId, opts?) → BoundAgent
Create a bound agent for identity and session lineage.
```
const agent = ai.agent('shopping-agent', {
  description?: string,   // human-readable, appears in event streams
  userId?: string,
  parentAgentId?: string,
  customerOrgId?: string,
  agentVersion?: string,
  env?: string,
  context?: Record<string, unknown>,
  sessionId?: string,
  groups?: Record<string, unknown>,
  deviceId?: string,
  browserSessionId?: string,
});
```

### agent.child(agentId, overrides?) → BoundAgent
Create a child agent that inherits parent identity.
```
const recipeAgent = shoppingAgent.child('recipe-agent', {
  description: 'Finds recipes and checks ingredient availability',
});
```

### agent.session(opts?) → Session
Create a session for multi-turn tracking.
```
const session = agent.session({
  sessionId?: string,        // defaults to random UUID
  userId?: string,
  deviceId?: string,
  browserSessionId?: string, // links to Amplitude browser session
  autoFlush?: boolean,       // auto-flush on completion (default: auto-detect serverless)
});
```

### session.run(fn) / session.runSync(fn)
Execute code within session context. Auto-ends session when fn completes.
Provider calls inside run() are automatically tagged with session/agent identity.
```
await session.run(async (s) => {
  s.trackUserMessage(content);
  // ... LLM calls are auto-tracked by provider wrappers ...
  s.trackAiMessage(response, model, provider, latencyMs);
});
// Session automatically ends here, [Agent] Session End emitted.
```

### s.runAs(childAgent, fn) / s.runAsSync(childAgent, fn)
Delegate to a child agent within the same session.
Shares sessionId, traceId, turn counter. Does NOT emit Session End for child.
```
await s.runAs(recipeAgent, async (cs) => {
  cs.trackUserMessage('Find me a pancake recipe');
  // LLM calls here are tagged with recipe-agent's agentId
  // parentAgentId is automatically set to shopping-agent
});
```

### s.trackUserMessage(content, opts?) → eventId
Track a user message. Returns the event ID.
```
s.trackUserMessage('What recipes do you have for pancakes?');
```

### s.trackAiMessage(content, model, provider, latencyMs, opts?) → eventId
Track an AI response. Use when provider wrappers can't auto-capture (e.g. Assistants API).
```
s.trackAiMessage('Here are some pancake recipes...', 'gpt-4o-mini', 'openai', 1250, {
  inputTokens: 150, outputTokens: 400, totalTokens: 550,
});
```

### s.trackToolCall(toolName, latencyMs, success, opts?) → eventId
Track a tool call explicitly (prefer tool() HOF instead).
```
s.trackToolCall('search_products', 45, true);
```

### s.score(name, value, targetId, opts?)
Track a quality score (user feedback, automated eval, reviewer).
```
s.score('user-feedback', 1, 'msg-001', { targetType: 'message', source: 'user' });
```

### ai.score(opts)
Track a score outside session context.
```
ai.score({ userId, name: 'user-feedback', value: 1, targetId: 'msg-001', targetType: 'message', source: 'user' });
```

### ai.flush() / agent.flush()
Flush pending events. CRITICAL for serverless (Next.js, Vercel, Lambda):
```
await session.run(async (s) => { /* ... */ });
await ai.flush(); // Must call before returning response in serverless!
```

### tool(fn, opts?) / tool(opts)(fn)
Higher-order function that wraps a function to auto-emit [Agent] Tool Call events.
Works inside session.run() — automatically inherits session/agent context.
```
import { tool } from '@amplitude/ai';
const searchProducts = tool(async (args: { query: string }) => { /* ... */ }, {
  name: 'search_products',
});
// Inside session.run(), just call the wrapped function:
const results = await searchProducts({ query: 'pancakes' });
// [Agent] Tool Call event is automatically emitted with duration, success, input/output.
```

### observe(fn, opts?) / observe(opts)(fn)
Higher-order function that wraps a function to auto-emit [Agent] Span events.
Creates a session boundary if none exists.
```
import { observe } from '@amplitude/ai';
const handleRequest = observe(async (req) => { /* ... */ }, { name: 'request-handler' });
```

### injectContext() / extractContext(headers)
Cross-service context propagation for message queues and microservices.
```
import { injectContext, extractContext } from '@amplitude/ai';
// Sender (inside session.run):
const headers = injectContext();
await queue.send({ payload, headers });
// Receiver:
const ctx = extractContext(message.headers);
const session = agent.session({ ...ctx });
```

### createAmplitudeAIMiddleware(opts)
Express/Fastify/Hono middleware for automatic session tracking.
```
import { createAmplitudeAIMiddleware } from '@amplitude/ai';
app.use(createAmplitudeAIMiddleware({
  amplitudeAI: ai,
  userIdResolver: (req) => req.headers['x-user-id'] ?? null,
}));
```

### MockAmplitudeAI (testing)
Deterministic test double. Captures events in-memory.
```
import { AIConfig } from '@amplitude/ai';
import { MockAmplitudeAI } from '@amplitude/ai/testing';
const mock = new MockAmplitudeAI(new AIConfig({ contentMode: 'full' }));
const agent = mock.agent('test-agent', { userId: 'u1' });
await agent.session({ sessionId: 's1' }).run(async (s) => { /* ... */ });
mock.getEvents()                              // all events
mock.getEvents('[Agent] Tool Call')            // filtered by type
mock.eventsForAgent('recipe-agent')           // filtered by agent ID
mock.assertEventTracked('[Agent] User Message', { userId: 'u1' })
mock.assertSessionClosed('s1')
await mock.flush();
```

---

## Provider Wrappers

| Provider    | Import                           | Streaming | Tool Calls | TTFB | Cache |
|-------------|----------------------------------|-----------|------------|------|-------|
| OpenAI      | `new OpenAI({ amplitude: ai })`  | Yes       | Yes        | Yes  | Yes   |
| Anthropic   | `new Anthropic({ amplitude: ai })`| Yes      | Yes        | Yes  | Yes   |
| Gemini      | `new Gemini({ amplitude: ai })`  | Yes       | No         | No   | No    |
| AzureOpenAI | `new AzureOpenAI({ amplitude: ai })`| Yes    | Yes        | Yes  | No    |
| Bedrock     | `new Bedrock({ amplitude: ai })` | Yes       | Yes        | No   | No    |
| Mistral     | `new Mistral({ amplitude: ai })` | Yes       | No         | No   | No    |

All wrappers are imported from '@amplitude/ai'.

---

## Canonical Patterns

### 1. Zero-code (patch)
```typescript
import { AmplitudeAI, patch } from '@amplitude/ai';
const ai = new AmplitudeAI({ apiKey: process.env.AMPLITUDE_AI_API_KEY! });
patch({ amplitudeAI: ai });
```

### 2. Wrap existing client
```typescript
import { AmplitudeAI, wrap } from '@amplitude/ai';
const ai = new AmplitudeAI({ apiKey: process.env.AMPLITUDE_AI_API_KEY! });
const instrumented = wrap(existingOpenAIClient, ai);
```

### 3. Provider wrapper (recommended)
```typescript
import { AmplitudeAI, AIConfig, OpenAI } from '@amplitude/ai';
const ai = new AmplitudeAI({
  apiKey: process.env.AMPLITUDE_AI_API_KEY!,
  config: new AIConfig({ contentMode: 'full', redactPii: true }),
});
const openai = new OpenAI({ amplitude: ai, apiKey: process.env.OPENAI_API_KEY! });
```

### 4. Multi-agent with session + runAs (recommended for multi-agent apps)
```typescript
import { AmplitudeAI, AIConfig, OpenAI, tool } from '@amplitude/ai';

const ai = new AmplitudeAI({ apiKey: process.env.AMPLITUDE_AI_API_KEY!, config: new AIConfig({ contentMode: 'full', redactPii: true }) });
const openai = new OpenAI({ amplitude: ai, apiKey: process.env.OPENAI_API_KEY! });

const orchestrator = ai.agent('shopping-agent', { description: 'Orchestrates shopping requests' });
const recipeAgent = orchestrator.child('recipe-agent', { description: 'Finds recipes' });

const askRecipeAgent = tool(async (args: { query: string }) => {
  // ... calls openai inside s.runAs(recipeAgent, ...)
}, { name: 'ask_recipe_agent' });

export async function POST(req: Request) {
  const { messages, userId } = await req.json();
  const browserSessionId = req.headers.get('x-amplitude-session-id');
  const deviceId = req.headers.get('x-amplitude-device-id');

  return orchestrator.session({ userId, browserSessionId, deviceId }).run(async (s) => {
    s.trackUserMessage(messages[messages.length - 1].content);

    // Provider calls auto-tracked. tool() calls auto-tracked.
    const response = await openai.chat.completions.create({ model: 'gpt-4o-mini', messages });

    s.trackAiMessage(response.choices[0].message.content, 'gpt-4o-mini', 'openai', latencyMs);
    await ai.flush(); // serverless: flush before returning
    return Response.json({ content: response.choices[0].message.content });
  });
}
```

### 5. Express middleware
```typescript
import { AmplitudeAI, createAmplitudeAIMiddleware } from '@amplitude/ai';
const ai = new AmplitudeAI({ apiKey: process.env.AMPLITUDE_AI_API_KEY! });
app.use(createAmplitudeAIMiddleware({
  amplitudeAI: ai,
  userIdResolver: (req) => req.headers['x-user-id'] ?? null,
}));
```

### 6. Testing with MockAmplitudeAI
```typescript
import { AIConfig, tool } from '@amplitude/ai';
import { MockAmplitudeAI } from '@amplitude/ai/testing';
const mock = new MockAmplitudeAI(new AIConfig({ contentMode: 'full' }));
const agent = mock.agent('test-agent');
const session = agent.session({ userId: 'test-user', sessionId: 'test-session' });
await session.run(async (s) => {
  s.trackUserMessage('hello');
});
expect(mock.getEvents('[Agent] User Message').length).toBe(1);
mock.assertSessionClosed('test-session');
```

---

## MCP Tools (via amplitude-ai mcp)

All tools are available when the MCP server is connected.

### scan_project(root_path) → ScanResult
Detect framework, providers, agents, call sites, and multi-agent signals.
Returns structured JSON with: framework, providers[], agents[] (each with inferred_id,
file, call_site_details, tool_definitions), is_multi_agent, multi_agent_signals[],
has_streaming, has_edge_runtime, has_frontend_deps, recommendations[].
Use this as the first step of instrumentation.

### validate_file(source, language?) → FileAnalysis
AST-based analysis of a single file for uninstrumented LLM call sites.
Returns: call_sites[] with instrumented (bool), has_patch, has_amplitude_import,
has_session_context. Run before and after instrumentation to verify coverage.

### instrument_file(source, language?, scan_result?) → InstrumentedSource
Apply text-based provider swap and session transforms.
Best for simple import swaps. For multi-agent patterns, prefer direct editing.

### generate_verify_test(scan_result) → TestSource
Generate a vitest verification test using MockAmplitudeAI.
Tests each discovered agent and multi-agent delegation.

### validate_setup() → SetupStatus
Check required environment variables (AMPLITUDE_AI_API_KEY, etc.).

### suggest_instrumentation(framework?, provider?, content_tier?) → Guidance
Framework-specific instrumentation guidance with content-tier and privacy defaults.

### get_event_schema(event_type?) → EventSchema
Return the event property catalog for all or specific event types.

### get_integration_pattern(id?) → Pattern
Return canonical instrumentation patterns (zero-code, wrap, bound-agent, etc.).

### search_docs(query, max_results?) → SearchResults
Keyword search over README and this file.

## MCP Resources

### amplitude-ai://event-schema
Full event property catalog (all [Agent] event types and their properties).

### amplitude-ai://integration-patterns
Canonical instrumentation patterns as structured JSON.

### amplitude-ai://instrument-guide
The complete 4-phase instrumentation workflow (Detect → Discover → Instrument → Verify)
with code examples for every step. Read this for guided instrumentation.

---

## CLI

- `amplitude-ai mcp` — Start the MCP server for AI coding agents
- `amplitude-ai doctor [--json] [--no-mock-check]` — Validate environment and event pipeline
- `amplitude-ai status [--json]` — Show SDK version, installed providers, and env config
- `amplitude-ai --help` / `amplitude-ai --version`

---

## Events Emitted

| Event Type | Emitted By |
|---|---|
| [Agent] User Message | s.trackUserMessage() |
| [Agent] AI Response | Provider wrappers (auto) or s.trackAiMessage() |
| [Agent] Tool Call | tool() HOF (auto) or s.trackToolCall() |
| [Agent] Span | observe() HOF (auto) or s.trackSpan() |
| [Agent] Session End | session.run() completion (auto) or s.trackSessionEnd() |
| [Agent] Score | ai.score() or s.score() |
| [Agent] Embedding | provider wrappers (auto) |
| [Agent] Session Enrichment | s.trackSessionEnrichment() |

---

## Common Errors

- "No events captured" → Ensure session.run() wraps your LLM calls
- "patch() drops events silently" → patch() requires active SessionContext; use session.run()
- "flush() timeout" → Call ai.flush() before process exit in serverless (Next.js, Lambda)
- "tool() not tracking" → tool() must be called inside session.run() to inherit context
- "child agent events missing parentAgentId" → Use s.runAs(child, fn), not direct child calls
- "observe() not emitting spans" → observe() must be called inside session.run() for session context
