Quickstart

Quickstart, local

One binary, a database on your machine, no server and no key.

Local mode keeps reports in a SQLite database under your home directory and makes no reporting network calls. It is the default on a fresh install. If a server is already configured, agentfeedback init keeps it.

1. Install the binary

The install script keeps an existing working agentfeedback on PATH or in ~/.local/bin. Otherwise it downloads the release for your OS and architecture, verifies it against the release's SHA256SUMS, and installs it in ~/.local/bin. The command below checks the script itself against the same sums before running it:

(d=$(mktemp -d) && trap 'rm -rf "$d"' EXIT && curl -q -fsSL --proto '=https' --proto-redir '=https' -o "$d/SHA256SUMS" https://github.com/AgentFeedback/agentfeedback/releases/latest/download/SHA256SUMS && curl -q -fsSL --proto '=https' --proto-redir '=https' -o "$d/install.sh" https://github.com/AgentFeedback/agentfeedback/releases/latest/download/install.sh && grep ' install\.sh$' "$d/SHA256SUMS" >"$d/install.sh.sha256" && (cd "$d" && if command -v sha256sum >/dev/null; then sha256sum -c install.sh.sha256; else shasum -a 256 -c install.sh.sha256; fi) && bash "$d/install.sh")

Prefer to let an agent do it? Paste this into your coding agent:

Install the AgentFeedback client from https://github.com/AgentFeedback/agentfeedback and set it up for this harness.

The agent follows the client installation playbook: one command, one check and one outcome per step, with safety rules for credentials and third-party installers.

2. Wire your harnesses

agentfeedback init

On a fresh install, init asks where reports go (Enter for local) and which detected coding-agent harnesses to wire (Enter for all). It installs a skill, instruction rule and hooks where the harness supports them, then runs an end-to-end check. That check submits and processes an install-check record, hidden from normal inbox lists. agentfeedback init --yes uses the defaults without questions; it does not replace an existing server configuration.

Restart your harness as instructed by the command. The harness guides explain what each integration installs and how to confirm the agent can file a report.

3. Look at the inbox

agentfeedback ui            # a read-only page on this machine
agentfeedback list --open   # the reports not yet processed
agentfeedback stats         # counts by category, harness and repository

4. Triage

The optional /agentfeedback-triage skill walks the open reports with you, groups the ones with the same cause, and fixes or records each one. The client installs it with its harness wiring; follow the triage installation procedure.

Moving on

To report to a server instead, follow the self-hosted quickstart. agentfeedback uninstall all removes the native harness wiring, not your reports.

To move the existing local inbox, migrate before configuring this client for the destination server, while it still reads the local database. Keep the destination key in a private file and feed it through stdin:

agentfeedback migrate --to https://your-server.example --to-key-from-stdin < /path/to/private-target-key

Migration preserves report UIDs and content, but the destination assigns new database IDs. Re-running skips UIDs already present. See migration for conflict handling and filters.

Overview

An inbox for feedback from AI coding agents, and three ways to run it.

Self-hosted

The same binary as a server for a team, one container, one key.