For AI agents: a documentation index is available at /docs/llms.txt. Append .md to any page URL for markdown, or send Accept: text/markdown.
Import agent conversations from other tools
Agent Analytics integrations bring conversations into Amplitude from where they already live. Hosted platforms, tracing tools, warehouses, and agent frameworks don't require the Amplitude AI SDK in your agent code. Gateways, Fireworks, and local models do. Each path turns conversations into the same [Agent] events the AI SDK sends, so sessions, built-in signals, evaluators, and charts work the same way.
Use an integration when your agent runs on a hosted platform, when a tracing tool already records it, when its conversations land in your warehouse, when the Cursor Agent SDK runs the model loop, or when an agent framework already exports OpenTelemetry. Use the Amplitude AI SDK when you instrument the model calls yourself. The SDK captures the most detail on that path, including cost, tool inputs, and user feedback. Refer to Instrument your agent with the AI SDK.
Set up with your AI coding agent
Amplitude writes each platform, tracing-tool, warehouse, and agent-framework guide for coding agents as well as engineers. Each one has a "Part 2: Coding agent procedure" that an agent follows from start to finish. Give your agent the page for your source from the table in Choose an integration, or point it at the full list:
- Guides index: https://raw.githubusercontent.com/amplitude/Amplitude-AI-Node/main/docs/integrations/README.md
- Machine-readable list: https://raw.githubusercontent.com/amplitude/Amplitude-AI-Node/main/docs/integrations/manifest.json
The GitHub guides are public and work with no SDK installed, and they track main. They also ship in the @amplitude/ai npm package. If it's installed, the same files are under node_modules/@amplitude/ai/docs/integrations/.
How integrations work
Integrations move conversations in one of five ways.
- Hosted platforms and tracing tools: A job in your infrastructure reads finished conversations, maps them to
[Agent]events, and sends them to the Amplitude HTTP API. Sierra can also push. - Agent frameworks: The framework exports OpenTelemetry. You don't run an Amplitude job.
- Cursor Agent SDK: A hook in the agent posts each live turn.
- Data warehouses: Amplitude runs your SQL on the schedule you set for table rows. For spans already stored in a table, your job replays them to the OTLP endpoint.
- Gateways, Fireworks, and local models: The Amplitude AI SDK records the calls. Fireworks and some gateways can also export OpenTelemetry.
Each integration has a step-by-step guide on GitHub, written for engineers and AI coding agents. Hand a coding agent the guide URL and it can build the integration. The docs pages here cover what each path captures and what to decide before you start.
Choose an integration
Hosted agent platforms
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| Sierra | Sierra | Sierra pushes, or your job pulls exports | Model, tokens, and cost |
| Decagon | Decagon | Your scheduled job pulls exports | Model, tokens, cost, and tool calls |
| Fin by Intercom | Fin | Your scheduled job pulls conversations | Model, tokens, cost, and tool inputs and outputs |
| Salesforce Agentforce | Agentforce | Your scheduled job queries Data Cloud | Cost |
Agent runtimes
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| Cursor Agent SDK | Cursor Agent SDK | A hook in the agent posts each live turn | Per-turn cost |
LLM tracing tools
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| Langfuse | Langfuse | Your scheduled job pulls sessions | Langfuse scores |
| LangSmith | LangSmith | Your scheduled job pulls threads | LangSmith feedback scores |
| Braintrust | Braintrust | Your scheduled job runs SQL on project logs | Cost and Braintrust scores |
Agent frameworks
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| AWS Strands (Python) | AWS Strands | Strands exports native OTLP/HTTP | [Agent] Session End, unless you also send one over HTTP |
| Google ADK | Google ADK | ADK exports native OTLP/HTTP | [Agent] Session End, unless you also send one over HTTP |
| Vercel AI SDK | Vercel AI SDK | The Vercel AI SDK exports native OTLP/HTTP | [Agent] Session End, unless you also send one over HTTP |
| Pydantic AI | Pydantic AI | Pydantic AI exports native OTLP/HTTP | [Agent] Session End, unless you also send one over HTTP |
| Mastra | Mastra | Mastra exports native OTLP/HTTP | [Agent] Session End, unless you also send one over HTTP |
| Microsoft Agent Framework | Microsoft Agent Framework | The framework exports native OTLP/HTTP | [Agent] Session End, unless you also send one over HTTP |
| OpenInference | OpenInference | Agno, smolagents, DSPy, Haystack, the Google GenAI SDK, and AutoGen | [Agent] Session End, unless you also send one over HTTP |
| Amazon Bedrock AgentCore | Amazon Bedrock AgentCore | Add OTLP/HTTP beside CloudWatch | [Agent] Session End, unless you also send one over HTTP |
TypeScript and Java name the model span call_llm until the exporter in the guide rewrites it. Go puts message text on OpenTelemetry logs, which Amplitude doesn't read. If your agent already emits OpenTelemetry GenAI spans and isn't listed here, add Amplitude as a second OTLP exporter. Refer to Send OpenTelemetry traces to Agent Analytics.
Inference routers
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| OpenRouter, LiteLLM, or Requesty | AI gateways and routers | The AI SDK, or the gateway's OTLP export | Depends on the path |
| Fireworks AI | Fireworks AI | The AI SDK, or your OpenTelemetry export | Nothing with the AI SDK. An OpenTelemetry export loses [Agent] Session End, unless you also send one over HTTP |
Data warehouses
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| Snowflake, BigQuery, or Databricks tables | Warehouse import | Amplitude imports with your SQL on your schedule | Anything your tables don't record |
| OpenTelemetry, OpenInference, or MLflow spans stored in a table | Warehouse import | Your job replays spans to the OTLP endpoint | Anything your spans don't record |
Local models
| Source | Integration | Who moves the data | What you lose compared with the AI SDK |
|---|---|---|---|
| Ollama, LM Studio, or another local model | Local models | The AI SDK with manual tracking | Cost |
Decide before you build
Every integration needs these values in the source data. The guides stop and ask rather than guess them.
- Conversation ID: The ID that groups messages into one session. The job skips and counts conversations without one, rather than sending them.
- User ID: The same user ID you send in product analytics, so agent sessions join your product data. The job skips conversations without one.
Verify an integration
Each guide starts with a dry run that builds events without sending them. The check-agent-events.mjs script ships in the @amplitude/ai npm package and checks a sample against the event contract. This install doesn't add the SDK to your app. Run it from the package rather than downloading it: npm install --no-save --ignore-scripts @amplitude/ai && node node_modules/@amplitude/ai/docs/integrations/check-agent-events.mjs events.json. After the first real run, check Live Events and open a session in Agent Analytics, then run the same window again to confirm that it adds no duplicates. Refer to Verify your Agent Analytics instrumentation.
Integration guides were last verified on October 6, 2026. The Strands guide was last verified on October 14, 2026. The Google ADK, Vercel AI SDK, Pydantic AI, Mastra, Microsoft Agent Framework, and OpenInference guides were last verified on October 28, 2026. The Amazon Bedrock AgentCore guide was last verified on October 27, 2026.
Was this helpful?