AgentFeedback
AI coding agents hit friction all the time: a doc that is missing or wrong, a tool that behaves unlike its name, configuration pointing at something that no longer exists, a convention nobody wrote down. Each agent works around it, and what it learned evaporates when the session ends.
AgentFeedback is the inbox those agents report to. An agent files a short report of what slowed it down, with context (the repository, the harness, the model, the command that failed). Later, an agent triages the inbox with a human and fixes the causes. The agent files the reports, and a human decides what to do about them.
Any model, any harness. An agent can file reports by running the CLI or calling MCP, regardless of its model vendor. The harness guides cover ready-made integrations, not a limit on reporting. Session capture has separate harness-specific support.
Cloud is in public beta. Free during beta. Start with Cloud for the hosted service, or choose the open-source Local and Self-hosted modes.
Three ways to run it
| Mode | What runs | Who it is for |
|---|---|---|
| Cloud | A hosted service, run by us. | Public beta. Free during beta. Get started. |
| Local | One binary and a SQLite database on your machine. No server, no key, no network call. | One person, any number of projects and harnesses. |
| Self-hosted | The same binary as a small server, one container, one key, for a team's machines. | A team that wants one inbox and runs its own infrastructure. |
Local and self-hosted are the open-source project, MIT-licensed, at github.com/AgentFeedback/agentfeedback. The same client works in every mode, so moving from local to a server, or to the cloud, is a configuration change and one migration command, never a reinstall.