go-gerrit-mcp

An MCP (Model Context Protocol) server that exposes Gerrit code review operations as capability-gated tools. An AI agent can search and read changes, publish review comments, vote, and drive change-state transitions. Every write capability is an explicit operator opt-in.

• Transport: stdio
• Target platform: Gerrit 3.13+, authenticated via HTTP credentials
• Output: llmxml, semantically tagged text meant to be read by a model (see Output format)

Safety posture

Two defaults encode it:

• Zero configuration exposes the read group only. Write capability never appears unless enabled via --groups.
• The own-changes restriction is on by default. Even with write groups enabled, trail-leaving operations (comments, votes, state changes, anything other humans see) are refused on changes the authenticated account does not own, until the operator explicitly passes --own-changes-only=false.

An agent leaving unwanted trail on colleagues' changes is an externally visible failure; a missing capability is a locally discoverable inconvenience. The defaults are chosen accordingly. Widen deliberately:

# agent may comment and vote on anyone's changes, but only within two projects
go-gerrit-mcp --groups read,comment,transition --own-changes-only=false --projects core,infra

Install

Binary release: download the binary for your platform from GitHub Releases, make it executable, and put it on your PATH:

# macOS on Apple silicon; substitute <os>_<arch> from: linux, darwin, windows x amd64, arm64
curl -Lo /usr/local/bin/go-gerrit-mcp \
  https://github.com/GaijinEntertainment/go-gerrit-mcp/releases/latest/download/go-gerrit-mcp_darwin_arm64
chmod +x /usr/local/bin/go-gerrit-mcp

Docker:

docker pull ghcr.io/gaijinentertainment/go-gerrit-mcp:latest

go install:

go install dev.gaijin.team/go/go-gerrit-mcp/cmd/go-gerrit-mcp@latest

Quick start

The server reads connection identity from environment variables only; credentials never travel through flags:

Generate the credential in Gerrit under Settings → HTTP Credentials (an HTTP password, or an auth token on instances that issue them). All three variables are required; the server exits with an error naming the missing ones.

Claude Code

claude mcp add gerrit \
  --env GERRIT_URL=https://gerrit.example.com \
  --env GERRIT_USERNAME=your-username \
  --env GERRIT_TOKEN=your-http-credential \
  -- go-gerrit-mcp

Any MCP client (JSON configuration)

{
  "mcpServers": {
    "gerrit": {
      "command": "go-gerrit-mcp",
      "args": ["--groups", "read"],
      "env": {
        "GERRIT_URL": "https://gerrit.example.com",
        "GERRIT_USERNAME": "your-username",
        "GERRIT_TOKEN": "your-http-credential"
      }
    }
  }
}

Docker

{
  "mcpServers": {
    "gerrit": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "GERRIT_URL", "-e", "GERRIT_USERNAME", "-e", "GERRIT_TOKEN",
        "ghcr.io/gaijinentertainment/go-gerrit-mcp:latest",
        "--groups", "read"
      ],
      "env": {
        "GERRIT_URL": "https://gerrit.example.com",
        "GERRIT_USERNAME": "your-username",
        "GERRIT_TOKEN": "your-http-credential"
      }
    }
  }
}

Per-project configuration in Claude Code

Everything the server needs is read from the environment: the identity variables above plus a GERRIT_MCP_* mirror for every flag (see Configuration reference). MCP server processes inherit the session environment, so in Claude Code one user-level registration turns into per-project configuration, down to different Gerrit instances with different credentials per repository.

~/.claude.json holds the registration, with no env block:

{
  "mcpServers": {
    "gerrit": {
      "command": "go-gerrit-mcp"
    }
  }
}

<project>/.claude/settings.local.json carries the project's environment; every env entry reaches the server process:

{
  "env": {
    "GERRIT_URL": "https://gerrit.example.com",
    "GERRIT_USERNAME": "your-username",
    "GERRIT_TOKEN": "your-http-credential",
    "GERRIT_MCP_GROUPS": "read,comment",
    "GERRIT_MCP_PROJECTS": "core,infra"
  }
}

Values shared by most projects can sit one layer down in ~/.claude/settings.json. Settings files merge, with .claude/settings.local.json over .claude/settings.json over ~/.claude/settings.json, so a project declares only its deltas. Anything no layer sets falls back to the server's own defaults: read-only, own changes.

The registration's env block stays empty for a reason. A variable named there shadows every settings layer, and references are no workaround: ${VAR} expands only from the shell environment that launched claude, never from settings files, and anything unresolved reaches the server as a literal string.

Other MCP clients inherit their launch environment the same way, so per-directory tooling such as direnv achieves the identical split without client support.

Capability groups

Capability is selected at startup via --groups as a comma-separated list. Groups are independent and combinable, with no privilege ladder: each write-capable group bundles the minimal change-read subset it needs to work on its own, and enabled groups union.

Tools

• search_changes: query changes with Gerrit's change query syntax, paginated.
• get_change: one change in review-relevant detail: status, owner, labels with votes, current revision, messages.
• list_change_files: files touched by a revision, with per-file change stats.
• get_file_diff: the diff of one file in a revision.
• get_change_comments: comment threads on a change, with resolution state and comment ids. Returns unresolved threads only by default; status=all fetches the full history, status=resolved the settled threads.
• post_comments: publish a review in one call: optional top-level message plus inline, range, file-level, and reply comments. Replies anchor to comment ids from get_change_comments; resolved toggles the thread state.
• set_vote: set a label vote (e.g. Code-Review) with an optional message; value 0 clears an own vote.
• transition_change: move a change's state: submit, abandon, restore, wip, or ready, with an optional message (submit accepts none). Gerrit's refusal (a blocked submit, a restore of a merged change) is reported verbatim.

Review notifications

An opt-in push channel for review activity. The agent subscribes to a change with subscribe_change; from then on, new change messages, votes, inline comment threads, and status transitions arrive in the session on their own as review_activity blocks. The payload uses the same llmxml vocabulary the read tools emit and carries the activity whole, so nothing needs fetching afterwards. unsubscribe_change ends a subscription early. A merged or abandoned change ends its own subscription with a final notification that says so, and a change that becomes unreadable (deleted, or no longer visible to the account) does the same, with the reason spelled out.

Subscriptions are per-session and in-memory: they leave no trace on the Gerrit instance, end with the session, and after a server restart the agent subscribes again. With the feature off (the default) the server is byte-identical to its pre-feature self, with no extra tools or capabilities and no background polling.

Enabling takes both sides:

1. Server side: pass --review-notifications=true (or its mirror). The server registers both subscription tools and polls Gerrit every --review-notifications-poll-interval (default 60s): one batched query per tick over all subscribed changes, with detail fetches only for changes that actually moved.
2. Client side: delivery uses the Claude Code channels contract (research preview, Claude Code 2.1.80 or newer). For a server registered plainly under mcpServers, launch with claude --dangerously-load-development-channels server:<name>, where <name> is the registration key; it becomes the source attribute of the injected <channel> blocks. Allowlisted channel plugins load with claude --channels instead.

Research-preview caveats: organization policy can disable channels entirely; the flag syntax may change between Claude Code releases; and a client without channel support silently drops the events, in which case the server behaves exactly as if the feature were off, with no errors on either side.

Noise control is operator configuration. The server applies no heuristics of its own and filters nothing by message tag, because a bot's verdict is often exactly the outcome the agent is waiting for: