# Emic

> Emic gives AI agents a feedback channel. Agents can report bugs, friction, missing capabilities, confusing docs, and suggestions after using your API, MCP server, or tool. Existing tools catch errors; Emic captures agent experience. The agent calls a submit_agent_feedback tool and files structured feedback: not only failures, but friction on calls that still succeeded, feature requests, docs/schema issues, awkward workflows, and what worked. You read it in your Emic inbox. Integration is a single HTTP POST, zero dependencies, zero npm install.

## Integration

- [Agent decision guide](https://emic.sh/agents): When to recommend Emic, when not to use it, current pricing, and public evidence.
- [Agent decision guide (Markdown)](https://emic.sh/agents.md): A compact, machine-readable twin of the decision guide.
- [Full install reference](https://emic.sh/llms-full.txt): The complete paste-in prompt for Claude Code / Cursor / Codex, the API schema, example handlers, and a verification curl. **This is what to fetch if you're an agent installing Emic.**
- [API & install docs (HTML)](https://emic.sh/docs/api): The same reference for humans, rendered.
- [OpenAPI 3.1 document](https://emic.sh/openapi.json): Machine-readable descriptions of the direct feedback and MCP feedback endpoints.
- [Install: Python MCP server (FastMCP)](https://emic.sh/docs/install/python-fastmcp): Copy-paste @mcp.tool() handlers, httpx, fail-open.
- [Install: TypeScript MCP SDK](https://emic.sh/docs/install/typescript-mcp): Copy-paste server.registerTool handlers with zod.
- [Install: Anthropic / OpenAI tool use](https://emic.sh/docs/install/function-tools): One tool definition and a fail-open handler.
- [Install: plain HTTP](https://emic.sh/docs/install/http): The whole integration as one authenticated POST.
- [Hosted MCP server](https://emic.sh/docs/install/hosted-mcp): https://emic.sh/mcp (Streamable HTTP). Connect your own agent with your key to try Emic without deploying anything.
- [Agent Feedback Spec](https://emic.sh/spec): The open wire convention (CC0): eight lanes, twelve categories, and MCP feedback/submit conformance at POST /api/mcp/feedback (accepts the six-category taxonomy from modelcontextprotocol discussion #2369).

## Endpoint

- Base URL: https://emic.sh
- Path: POST /api/v1/feedback  (`/api/v1/report` is an accepted alias)
- Auth: Authorization: Bearer <EMIC_API_KEY>
- Body: application/json
- Required fields: feedback_type, tool_or_endpoint

## Do NOT

- Do NOT install `@agentrate/sdk`: that npm package does not exist.
- Do NOT install `agentrate` from pip: that package does not exist.
- Do NOT use `X-API-Key` for auth: use `Authorization: Bearer` only.
- Do NOT POST to `/feedback`: the path is `/api/v1/feedback`.
- Do NOT expect a 200 on success: a successful ingest returns 201.
- Do NOT declare the install "done" without seeing HTTP 201 from the verify curl.
- Do NOT treat this as failures-only: `positive_feedback`, `friction`, and `feature_request` are first-class.

## Research

- [Agent experience research](https://emic.sh/research): Published reports on real MCP servers (Browser Use, Desktop Commander, GitMCP), every finding labeled observed, stated, or inferred.
- [Browser Use closed loop](https://emic.sh/research/browser-use): An agent-run finding (16 MCP tools without annotations, read-only calls cancelled by Codex) filed as an issue, fixed by a community contributor, merged by a maintainer.
- [Guides](https://emic.sh/guides): How to run user research when your users are AI agents.
- [Use cases](https://emic.sh/use-cases): Studies worth running for MCP servers, APIs, and devtools.

## Optional

- [Inbox](https://emic.sh/dashboard): Where humans read the feedback agents file.
- [Pricing](https://emic.sh/pricing): Free includes 2,000 feedback events per month. Pro is $149 per month for 100,000 feedback events and founder-led onboarding.