gerrit-mcp-server

한국어 문서

Connect your AI coding agent to Gerrit code review.

Ask your agent to find the changes waiting on you, read a diff, draft line comments, and publish a review — without leaving the session and without pasting code review comments back and forth.

It ships as two frontends over the same code, so you can pick how much of your agent's context you want to spend:

The skill route is the lighter one and works with any agent that reads skills. The MCP server needs no shell access and works with any MCP client: Claude Code, Codex, Cursor, Zed, Continue, or your own.

Either way it is a single static binary with no runtime dependencies. You self-host it; nothing is sent anywhere except to the Gerrit host you configure.

Quick start

Both routes start the same way.

Create a Gerrit auth token. In Gerrit, go to Settings → HTTP Credentials and generate one. See Credentials below if your Gerrit is older.

Node is needed only so that npm or npx can fetch the right binary for your machine. The binaries are Go and have no Node runtime dependency.

Route A — gerrit-cli + agent skill

The lighter one. Nothing sits in your agent's context until it needs Gerrit.

1. Install the binary.

npm i -g @gyeonghokim/gerrit-cli

2. Install the skill. This works for Claude Code, Codex, Cursor and many others.

npx skills add GyeongHoKim/gerrit-mcp-server

3. Configure it. Run this yourself in a terminal — it asks for your token on stdin, and will refuse to run where nothing can type into it.

gerrit-cli init

4. Check it. gerrit-cli config reports every setting and where it came from, naming anything still missing. Then ask your agent: "What’s the verified score for XXX Change Id’s review?"

To allow the commands that modify Gerrit, set GERRIT_ALLOW_WRITE=true — see Available tools and commands.

You can also use the CLI on its own, without an agent:

gerrit-cli query-changes --query "is:open reviewer:self -owner:self"
gerrit-cli get-file-diff --change-id 12345 --file src/main.go
gerrit-cli help

Route B — MCP server

No shell access needed, and it works with any MCP client.

Add the server to your client.

claude mcp add gerrit \
  --env GERRIT_URL=https://gerrit.example.com \
  --env GERRIT_USER=your-username \
  --env GERRIT_TOKEN=your-token \
  -- npx -y @gyeonghokim/gerrit-mcp-server

Add this to ~/.codex/config.toml, or to .codex/config.toml for a single trusted project:

[mcp_servers.gerrit]
command = "npx"
args = ["-y", "@gyeonghokim/gerrit-mcp-server"]
# npx downloads the binary on first run, which can exceed the 10s default.
startup_timeout_sec = 60

[mcp_servers.gerrit.env]
GERRIT_URL = "https://gerrit.example.com"
GERRIT_USER = "your-username"
GERRIT_TOKEN = "your-token"

Or let the CLI write it for you:

codex mcp add gerrit \
  --env GERRIT_URL=https://gerrit.example.com \
  --env GERRIT_USER=your-username \
  --env GERRIT_TOKEN=your-token \
  -- npx -y @gyeonghokim/gerrit-mcp-server

Add this to the client's MCP configuration file:

{
  "mcpServers": {
    "gerrit": {
      "command": "npx",
      "args": ["-y", "@gyeonghokim/gerrit-mcp-server"],
      "env": {
        "GERRIT_URL": "https://gerrit.example.com",
        "GERRIT_USER": "your-username",
        "GERRIT_TOKEN": "your-token"
      }
    }
  }
}

Ask for something. "What changes am I reviewing?" or "Summarise the diff on change 12345."

Related MCP server: mcp-gerrit-code-review

Credentials

Gerrit authenticates REST clients with HTTP Basic using a token from your account settings, and expects authenticated requests to be prefixed with /a/. Both binaries handle the prefix for you.

Generate a token under Settings → HTTP Credentials. On Gerrit 3.13 and newer you can name the token and give it a lifetime (90 days, 1 year, and so on) — worth doing, so the credential this holds is scoped and expires on its own. Older Gerrit versions call the same thing an HTTP password; it still works, as that endpoint is now an alias that creates a token with the id legacy.

Where the token lives depends on which frontend you use.

gerrit-mcp-server reads the environment and only the environment, so its credentials live in your MCP client's config file and nowhere else. It never reads a file of its own.

gerrit-cli has no client config to inherit from, so gerrit-cli init writes one under the OS configuration directory:

Set GERRIT_CONFIG to put it somewhere else. Environment variables always take precedence over the file, so a one-off GERRIT_TOKEN=... gerrit-cli ... works and CI never needs a file at all. gerrit-cli config prints where each value actually came from.

What holds for both:

• The token only ever travels in an Authorization header. It is never passed as a process argument, so it cannot be read out of ps, and it is never written to a log line or an error message. gerrit-cli init has no --token flag for exactly this reason.

• Nothing goes anywhere but your Gerrit host.

What is worth knowing about the file:

• On Linux and macOS it is written 0600, readable only by you. On Windows it inherits the ACL of %AppData%, which is already restricted to your account plus SYSTEM and Administrators — setting a tighter one needs a dependency this project does not take. If your %AppData% is redirected to a network share, prefer keeping GERRIT_TOKEN in your environment instead.

• gerrit-cli init echoes the token as you type it. Hiding terminal input needs a dependency this project does not take either. Pipe it in if that matters: printf 'https://gerrit.example.com alice %s ' "$TOKEN" | gerrit-cli init -non-interactive

• Do not commit it, and do not commit an MCP config either. A project-level .mcp.json holding GERRIT_TOKEN is a credential in your repository. Keep it in your user-level client config, or gitignore it.

• Use a dedicated token with a lifetime, so it can be revoked without touching your other credentials.

• Your Gerrit permissions still apply. Neither frontend can see or do anything your account cannot.

Configuration

Both frontends read the same variables. For the MCP server they live in your client's config; for the CLI they are optional, since gerrit-cli init writes the same settings to a file.

Available tools and commands

The two frontends expose exactly the same set, and a test in the repository holds them there. A CLI command is its MCP tool name with the underscores written as dashes — query_changes becomes query-changes — and gerrit-cli accepts either spelling.

Reads are always available. Writes are off unless you set GERRIT_ALLOW_WRITE=true, so an agent cannot abandon a change or post a review by accident. The MCP server does not register the write tools at all; gerrit-cli still lists them in its help, marked, but refuses to run one.

The same asymmetry applies to operations your Gerrit is too old for — see Supported Gerrit versions.

Read

Every value is a flag; gerrit-cli has no positional arguments. Run gerrit-cli help <command> for one command's flags — that is authoritative and cannot go stale.

gerrit-cli also has five commands of its own that no MCP tool corresponds to: help, version, config, init, and doctor, which reports your host's Gerrit release and what that rules out.

There is deliberately no --json output. Everything passes through the same renderer that keeps responses inside a sensible token budget, and handing an agent raw Gerrit JSON would undo that.

Write — requires GERRIT_ALLOW_WRITE=true

Exit codes

gerrit-cli reports what to do about a failure, not just that one happened. Rendered output goes to stdout and everything else to stderr, so the answer is safe to pipe.

Supported Gerrit versions

Built and tested against the Gerrit 3.14 REST API. Supported down to 2.14.

Almost everything works unchanged on an old host — including the draft comment endpoints, which earlier versions of this document blamed. Three write operations genuinely do not exist:

The two frontends handle that the same way they handle write access. gerrit-mcp-server asks the host which release it is as it starts, and never offers a tool it cannot serve, so the tool list your client sees is right from the first time it asks. gerrit-cli lists the commands with the release each needs and reports exit 4 if you run one anyway, naming both the release required and the one your host reports.

If the release cannot be determined, everything is offered. A proxy that swallows the version endpoint, or a patched internal fork that backported an endpoint, should not lose an operation that works — so an unknown version hides nothing, and anything genuinely missing still fails with a clear message.

One more difference worth knowing: Gerrit did not report comment counts before 3.0, so get_change_details omits that line on an older host rather than claiming zero.

gerrit-cli doctor    # which release your host runs, and what that rules out

Other ways to install

npm is the easy path, but the binaries stand alone.

# Go toolchain
go install github.com/GyeongHoKim/gerrit-mcp-server/cmd/gerrit-mcp-server@latest
go install github.com/GyeongHoKim/gerrit-mcp-server/cmd/gerrit-cli@latest

Or download the archive for your platform from the releases page. It contains both binaries and the agent skill, and you can point your MCP client's command straight at gerrit-mcp-server. No Node required.

Development

mise install     # toolchain, pinned in mise.toml
just setup       # dependencies and git hooks
just ci          # everything CI runs
just --list      # all tasks

If you are AI Coding Agent(Codex, Claude Code, OpenCode, etc.), See AGENTS.md for architecture, conventions, and the Gerrit API details worth knowing before you touch the client.

License

Elastic License 2.0.

It is prohibited to deploy this MCP server as a commercial service.

Using this at work is fine. ELv2 places exactly three restrictions on you: you may not offer this software to third parties as a hosted or managed service, you may not circumvent license key functionality, and you may not strip the copyright notices. Running it, modifying it, forking it, and deploying it across your engineering organisation are all expressly permitted.

Note that ELv2 is source-available rather than OSI-approved open source. If your organisation screens dependencies by license, it may need to be allowlisted.