Agentback docs

Agentback collects feedback from the AI agents that use your product, and from the people in your web app, into one inbox. An agent that gets blocked by your MCP server, CLI or API tells you what it was trying to do, what went wrong and what would have helped. You read the reports in the inbox, get them by email, or hand them to your coding agent to triage. It is a direct way to measure your product's agent experience.

How it works

Create a project. Each project has a public project key (af_...). It can only submit reports, so it is safe to ship inside an MCP server, a CLI or a web page.
Add Agentback where agents and people use your product. The agentback npm package adds a send_feedback tool to an MCP server, a feedback command to a CLI, and a feedback panel to a web app. Anything else sends one HTTP POST.
Agents report when they get stuck. The tool description tells agents when to report: when they are blocked, work around a problem, or finish only after retries or guessing.
Triage in one inbox. Reports arrive as new. Move them to in progress, resolved or ignored in the inbox, or let your coding agent do it over Agentback's MCP server.

Pick an integration

Where your product is used	Guide	What agents or people get
MCP server	Add a feedback tool to your MCP server	A `send_feedback` tool next to your own tools
CLI	Add a feedback command to your CLI	A `feedback` command and a hint on errors
HTTP API, any language	Collect agent feedback over HTTP	A public address that also explains itself to agents
Web app	Add a feedback widget to your web app	A Feedback button and a small panel

A product can have more than one surface. Use the same project for all of them, and each report records where it came from.

Let your coding agent do the setup

You can skip the manual steps. Paste this prompt into Claude Code, Cursor, Codex or another coding agent in your product's repository:

text
Add Agentback to this project so AI agents, and people using its web app, can report problems they run into.
Follow the guide at https://agentback.app/onboard.md.
If I haven't given you a project key, connect to my Agentback account as the guide describes.

The agent finds your MCP server, CLI, API or web app, starts a connection, and asks you to approve it in the browser. You sign up or sign in, pick or create the project, and the agent receives the project key. It never sees your secrets.

Keys

Key	Prefix	Where it may live	What it can do
Project key	`af_`	Anywhere, including code you ship	Submit reports
Secret key	`af_sk_`	Your backend or a hosted MCP server only	Submit reports with an account and request context
Identity secret	`af_id_`	Your backend only	Sign account IDs for verified accounts

Keep the secret key and the identity secret in environment variables, never in source code, prompts or anything that reaches end users. Rotating a project's keys replaces both.

What a report contains

Every report has a message. Agents also say how their task ended (blocked, worked around, or friction), what they were trying to do, what they expected and what happened instead, the tool involved, and what would have helped. Your integration adds the source, your product's version and, if you choose, a verified account. The full field list is in the HTTP guide.

Hand the setup to your coding agent

Paste one prompt into Claude Code, Cursor or Codex. It finds your MCP server, CLI, API or web app, connects to your account and adds Agentback.

Create a free inbox

MCP server