Google Chat MCP Server

google-chat-mcp

Google Chat as MCP tools. Read, search and write to your spaces, DMs and sidebar.

A single Go binary that speaks MCP over stdio. It runs as a subprocess of your client, on your own machine, against your own Google account. There is no server to host, no shared deployment and no service account: you create a Google OAuth client, log in once, and the refresh token stays in your OS keyring. Claude Code, Claude Desktop and any other MCP client can drive it.

It is for one person who uses Google Chat, on a Workspace or a personal Google account, and wants an assistant to read and write Chat as them. You create your own OAuth client in Google Cloud once; on Workspace, your administrator has to allow user-managed OAuth clients. A Workspace administrator can also turn on a toolset that manages every named space in the organization.

Goals and non-goals

Goals:

• Act as you. Every call runs as the signed-in account, with its access and nothing more.
• One binary per person. It runs on your machine, as a subprocess of your client, with your own OAuth client and token. Nothing is hosted or shared.
• Safe by default. What cannot be taken back is registered only when you turn it on, read-only mode removes every write, and most writes can be previewed with dry_run.
• Nothing added, nothing logged. send_message posts the body verbatim, and the logs never carry message text, addresses or search terms.
• Keep working when Google changes. An unknown field in a response is logged by name, never fatal.

Non-goals:

• A Chat app or bot. There is no service account, no app authentication and no domain-wide delegation, so it never posts as an app or acts for another user.
• A hosted service. It speaks stdio only, so Claude on the web and on mobile cannot reach it.
• Reading other people's messages as an administrator. The admin toolset manages spaces and memberships. Google offers no administrator access to messages.
• Filtering what a message says. Message text reaches your client verbatim, and this server cannot tell an instruction in it from a sentence.
• Asking you before a call. Your client approves each call; the server asks nothing.

Why google-chat-mcp

Chat integrations are normally apps: they post as a bot, see only the spaces somebody added them to, and need a service account an administrator provisions. This one runs as you. It reaches what your account already reaches, and messages you send are from you, not from an app with your name on it.

That also puts the user-scoped half of Chat in reach, which an app identity cannot get to at all: your sidebar sections, how far you have read in a space or a thread, your availability and custom status, and your organization's custom emoji. Login asks for the administrator scopes only when you turn the admin toolset on, and only for read-only scopes under GCM_READ_ONLY.

Install

go install github.com/mmedum/google-chat-mcp/v9/cmd/google-chat-mcp@latest

Or take a signed archive from the latest release — Linux, macOS and Windows, on amd64 and arm64 — and verify it before you run it:

sha256sum -c checksums.txt --ignore-missing

# The checksum file is signed with a keyless Sigstore certificate tied to
# the release workflow's identity. The bundle carries both.
cosign verify-blob checksums.txt \
  --bundle checksums.txt.bundle \
  --certificate-identity-regexp 'https://github\.com/mmedum/google-chat-mcp/' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

# And the archive itself carries build provenance.
gh attestation verify google-chat-mcp_*.tar.gz --repo mmedum/google-chat-mcp

Every archive also ships an SBOM, so you can see what is inside a binary you did not build.

Claude Desktop

Every release also carries a .mcpb bundle. Open it and Claude Desktop installs the server and asks for the settings a config file would carry, with no file to edit: your OAuth client JSON, which may be left empty to use the one login recorded, the profile, read-only, the toolsets and the attachment folder. The tools that cannot be taken back (enable_destructive) stay off unless you turn them on, and so does attachment transfer until you pick a folder. This server asks nobody before a call. Claude Desktop approves each tool, and a tool you always allow then runs unasked, so leave enable_destructive off unless you need its tools.

It covers macOS, Windows and Linux on both architectures each: macOS through a universal binary, Windows through amd64, and Linux through a small launcher that picks the right binary at start, because a bundle manifest has no key for the architecture. Its SHA-256 is in the same signed checksums.txt. The release is listed in the MCP registry as io.github.mmedum/google-chat-mcp.

The bundle does not log you in. Install the binary as well, run google-chat-mcp login once, and the bundle picks up the same credentials. Claude Code does not install .mcpb files, so it uses the command below.

Set up Google

You need your own OAuth client. It takes about fifteen minutes once, and docs/gcp-setup.md walks through it. In short: create a project, enable the Google Chat API, the People API and the Cloud Identity API, configure the consent screen, list the scopes this server asks for if the app is External, and create an OAuth 2.0 Client ID of type Desktop app. Download the JSON.

Then log in:

google-chat-mcp login --client-secret ./client_secret.json

Login opens your browser, or prints the URL when there is no browser to open (--no-browser forces that). The callback lands on 127.0.0.1 on a random port, with PKCE and state throughout. The refresh token goes into your OS keyring; if there is no keyring, it goes into a 0600 file and the command tells you so. On Windows a Go file mode only sets the read-only attribute, so what protects that file there is the ACL its directory inherits, not the mode.

google-chat-mcp status says which account is signed in and where the token lives. google-chat-mcp logout revokes the token at Google and deletes the local copy.

google-chat-mcp status --json prints the same state as one JSON object on stdout, for a script that needs to know whether this server is authorized before starting it. credentials.resolved is the field to branch on, credentials.signed_in distinguishes "no profile yet" from "profile but no token", schema_version changes only when a field is removed or its meaning changes, and the account is masked to its domain exactly as the text output masks it. A label in the human output is free to be reworded in any release; the object is not.

Logging in over SSH

The callback goes to the remote host's loopback address, and your browser is local, so forward the port. It is chosen at random and printed only once login is already waiting, so read it out of the printed URL — it appears percent-encoded, as 127.0.0.1%3A<port> — and in a second local terminal:

ssh -N -L <port>:127.0.0.1:<port> user@remote-host

Then open the URL locally. If ssh reports bind: Address already in use, cancel the login with Ctrl-C and run it again to draw another port.

Connect a client

Claude Code:

claude mcp add google-chat -- google-chat-mcp

Or, in a client config file:

{
  "mcpServers": {
    "google-chat": {
      "command": "google-chat-mcp"
    }
  }
}

Claude Desktop reads the same shape from its own config file. Use the binary's absolute path there if it is not on the app's PATH, or install the .mcpb bundle above and edit nothing.

Every setting is an environment variable, listed in docs/configuration.md. These are worth knowing first: GCM_READ_ONLY=true leaves out every tool that changes anything in Chat, GCM_ENABLE_DESTRUCTIVE=true adds the tools that cannot be taken back, GCM_PROFILE lets one machine hold a work account and a personal one, and GCM_LOCAL_DIR names the one directory files may be written to and read from. Without that last one the server touches no files at all, and the tools that move files say so.

Tools

Sections are your own sidebar. Nothing there changes a space, and nobody else sees it.

The admin toolset is for a Workspace administrator, and is off unless GCM_TOOLSETS names it, as in all,admin. Its tools act on any named space in the organization, spaces you are not in included, with your administrator privileges. None of them reads messages: Google offers no administrator access to those. docs/gcp-setup.md lists the scopes it adds.

These resources carry the same content as the matching tools, for a client that includes resources in its context:

• gchat://spaces/{space_id}
• gchat://spaces/{space_id}/messages/{message_id}
• gchat://spaces/{space_id}/threads/{thread_id}

Safety

• dry_run on 36 write tools. It returns the request body that would have been sent, and the call cannot write: the flag puts the request on a context the HTTP client refuses to write under, so a tool that forgot its own preview branch fails loudly instead of posting. A preview may still read what it shows or checks, such as the quoted message a quote names or the text a delete repeats.

• GCM_READ_ONLY=true leaves the write tools unregistered. A tool that is not registered cannot be called, whatever permission mode the client is in or whatever a model asks for. It is set where you start the server, so it binds the whole session rather than one call. A google-chat-mcp login run with it set asks for read-only scopes, so that token cannot write either. A token from an earlier login keeps its write scopes until you log in again, and turning writes back on takes another login.

• What cannot be taken back is not registered unless you say so. The deletes, delete_thread, set_space_audience, promote_member and add_external_member exist only with GCM_ENABLE_DESTRUCTIVE=true, and so do the admin tools that delete, promote, add someone or change a space's permissions. Each call then says confirm: true. delete_message, delete_thread, delete_space, admin_delete_space, delete_custom_emoji and set_space_audience also repeat what they act on: how the message begins, or the space's or the emoji's name. The server reads the target first and refuses [stale] when it differs, so the call your client shows you says which message or space goes.

• Your client approves each call; the server asks you nothing. Claude Code asks for the tools you list under permissions.ask. claude.ai and Codex prompt for a tool marked destructiveHint, which every tool in the list above that needs GCM_ENABLE_DESTRUCTIVE carries, and so does every write that replaces or removes something, such as update_message, update_space or remove_reaction. VS Code prompts for every tool that changes anything, download_attachment included. In Claude Code, put the writes that post, invite, share or delete under permissions.ask in your settings, with the name you gave the server in place of google-chat:

  {
    "permissions": {
      "ask": [
        "mcp__google-chat__send_message", "mcp__google-chat__update_message",
        "mcp__google-chat__upload_attachment",
        "mcp__google-chat__create_space", "mcp__google-chat__create_group_chat",
        "mcp__google-chat__find_direct_message",
        "mcp__google-chat__add_member", "mcp__google-chat__add_external_member",
        "mcp__google-chat__update_member_role", "mcp__google-chat__promote_member",
        "mcp__google-chat__remove_member", "mcp__google-chat__update_space",
        "mcp__google-chat__pin_message", "mcp__google-chat__create_custom_emoji",
        "mcp__google-chat__delete_message", "mcp__google-chat__delete_thread",
        "mcp__google-chat__delete_section",
        "mcp__google-chat__admin_update_space",
        "mcp__google-chat__admin_update_space_permissions",
        "mcp__google-chat__admin_update_member_role",
        "mcp__google-chat__admin_promote_member",
        "mcp__google-chat__admin_add_member",
        "mcp__google-chat__admin_remove_member"
      ]
    }
  }
  

  delete_space, admin_delete_space, set_space_audience and delete_custom_emoji need no rule: they carry Claude Code's requiresUserInteraction mark, so Claude Code 2.1.199 or later asks on every call in every interactive mode, and refuses them where nobody can be asked.

• Nothing reaches the whole space or past it by default. A post or an edit that mentions everyone, <users/all>, is refused. A space is created private, and update_space does not open one. In a space that admits people from outside your organization, add_member refuses an address your organization's directory does not know and every Google Group, and update_member_role makes nobody a manager: those are add_external_member and promote_member.

• send_message posts the body verbatim. No prefix, no suffix, no "sent by an assistant" footer.

• A retried post cannot land twice. Each message carries a client-chosen id, so a retry after a timeout returns what already landed instead of posting again.

• Logs never carry the payload. They record which call was made, to which resource, and how it ended. Message text, addresses, names and search terms stay out, and a test fails the build if one appears.

How it works

MCP client ──stdio──► google-chat-mcp
                       ├── tools      one handler per tool; shapes the reply
                       ├── service    the rules: idempotency, degrading, search
                       ├── gchat      REST client for Chat, People, Cloud Identity and OIDC
                       ├── directory  email cache, so a page of messages costs one lookup
                       └── auth       refresh token → access token

Every space, message and thread is addressed by its resource name, never by position, and a listing keeps paging while Google returns a token — including past an empty page, which Chat returns when it applies the page size before the filter.

Getting help

If something does not work, run google-chat-mcp status and google-chat-mcp doctor. status says which account is signed in, where the token lives and which scopes it lacks. doctor calls Google and names each check that could not run, and why. Most first-run trouble is an API that was never enabled or a scope that was never granted. docs/troubleshooting.md has the fixes.

If that does not explain it, open an issue. Never paste a space or message resource name, message text, a client_secret.json or a token into one; describe the shape instead. Security problems go through SECURITY.md, privately.

Versioning

Tool names and their output fields, and the resource templates, are stable within a major version. A change needing you to act — an added scope, another login, a different command in your client config — is marked Breaking: in CHANGELOG.md, which is also what each release's notes are made from.

Development

make build     # the binary
make test      # race detector, with a coverage profile
make cover     # the tests, then the coverage floor
make check     # the gates CI runs

make check is the definition of done. Among its gates, the race tests hold each package to an 80% coverage floor, except cmd/, which has a lower floor of its own. CONTRIBUTING.md describes the others.

Documentation

• docs/architecture.md — the request flow, the package layout, and the design decisions a contributor should not undo.
• docs/configuration.md — every environment variable.
• docs/gcp-setup.md — the Google Cloud project and OAuth client, step by step.
• docs/security.md — the threat model.
• docs/troubleshooting.md — symptoms and their fixes, then rotating the client secret, revoking access and running two accounts.
• docs/release.md — how a release is cut and checked.

Contributing

Questions and bugs go in issues; pull requests are welcome. CONTRIBUTING.md covers the branch and review flow, and make check is what has to pass.

Security

SECURITY.md says how to report a vulnerability.

Code signing policy

The Windows programs are not signed yet, so Windows Smart App Control may block them. This project is applying to SignPath Foundation, which signs open-source software for free, and signed releases will follow this policy.

Free code signing provided by SignPath.io, certificate by SignPath Foundation.

• What is signed: google-chat-mcp.exe for amd64 and arm64, in the Windows release zips and in the Claude Desktop bundle. Nothing else in a release is a Windows program.
• How it is built: the release workflow in this repository builds both programs from the release tag, on GitHub-hosted runners. SignPath signs only what that workflow built.
• Approval: an approver approves every signing request by hand.
• Committers and reviewers: mmedum
• Approvers: mmedum

Privacy: this program will not transfer any information to other networked systems unless specifically requested by the user or the person installing or operating it. It talks to Google alone: to sign you in, and to Google's APIs for the account you logged in with, when you run one of its commands or your MCP client calls one of its tools. Google's privacy policy covers what Google does with that.

Code of conduct

CODE_OF_CONDUCT.md — Contributor Covenant 3.0.

License

Apache 2.0 — see LICENSE.