google-mcp

A consolidated Google Workspace MCP server: Gmail, Calendar, Meet, Drive, Docs, Sheets, and Slides tools in a single stdio process. Written in TypeScript against the official googleapis client.

Process model: one process per account

Each running server is bound to exactly one Google account, selected with --slug:

google-mcp --slug jane-acme-com [--token-dir <dir>]

• Credentials are read from <token-dir>/google-<slug>-credentials.json; --token-dir defaults to ~/.config/openbrain/tokens, so the usual path is ~/.config/openbrain/tokens/google-<slug>-credentials.json.
• The credentials file holds the OAuth client id/secret and a refresh token. Refreshed access tokens are automatically persisted back to the file (atomic tmp-file + rename), so new processes reuse them instead of re-refreshing.
• In practice these processes are launched by openbrain's google-mcp.sh, one per configured account.

Adding an account

Accounts are provisioned by bootstrap/lib/add-google-account.sh, which lives in the external openbrain repo (not in this repository). It runs the OAuth consent flow and writes the credentials file for the slug. Re-run it for an existing slug to grant newly required scopes (error messages from this server reference it when a refresh token has expired or been revoked).

Hosted / env-token mode

For hosts that own token custody themselves (e.g. a hosted runtime that decrypts a per-connection OAuth bundle and refreshes tokens centrally), the server can instead take a ready-to-use access token from an environment variable:

GOOGLE_ACCESS_TOKEN=ya29... google-mcp --access-token-env            # reads GOOGLE_ACCESS_TOKEN
MY_TOKEN_VAR=ya29...       google-mcp --access-token-env MY_TOKEN_VAR

• The token is read once at startup and used for the life of the process. Refresh is the host's job — there is no refresh token, no OAuth client id/secret, and on a 401 the readable API error is surfaced for the host to classify; the server never attempts a refresh.
• Nothing touches disk: no credentials file is read and no token writeback of any kind happens.
• The flag is mutually exclusive with an explicit --token-dir. --slug becomes optional and, when given, is only used as an account label in error hints.
• Startup fails fast with a clear error naming the variable if it is unset or empty.

Tool families

| Family | Prefix | What it covers | |---|---|---| | Gmail | gmail_* | Search/read/send/reply (threaded, with attachments), threads, drafts, labels, filters, vacation responder, profile | | Calendar | calendar_* | List/create/update/respond to events, recurring-event instances, free/busy, propose meeting times, Meet links on events, out-of-office/working-location events | | Meet | meet_* | Create standalone Meet spaces, list conference records, fetch transcripts | | Drive | drive_* | List/search/move/copy/rename/delete, folder management, download/upload, sharing and permissions | | Docs | docs_* | Read/write Google Docs (native markdown in/out), text styling, tables, tabs, comments | | Sheets | sheets_* | Read/write values (single, batch, append), sheet/tab management, formatting, borders/merges, charts, find/replace, named ranges | | Slides | slides_* | Create/edit presentations, slides, text/shapes/images, styling, PDF export, thumbnails |

By default every family registers. --families gmail,calendar restricts the server to just the listed families (valid names: gmail, calendar, meet, drive, docs, sheets, slides); an unknown name fails startup with an error listing the valid ones. Works in both auth modes.

Operational defaults

• Compact output — tool results are unindented JSON; responses are pruned to the fields a model actually needs (list views trim heavyweight nested objects).
• Size caps — docs_read_document caps output at 50k characters (raise via maxLength); Gmail message bodies are capped at 50k characters with a truncation note; Sheets reads fail loudly past a 10,000-cell cap (narrow the range and retry); Gmail attachment downloads over 2 MB are written to disk instead of returned inline (Drive downloads inline up to 10 MB).
• Pagination — list-style tools accept a pageToken and return a nextPageToken when more results exist; tools that fan out to multiple underlying lists use a single composite token with the same contract.
• Retries — every Google API call retries 429s and 5xx errors up to 3 times globally, honoring the server's Retry-After header.
• Error mapping — all tool handlers are wrapped centrally; Google API failures come back as readable errors (HTTP status + API message) with hints for expired auth and quota problems.
• Token persistence — refreshed access tokens are written back to the per-account credentials file automatically.
• Downloads — files fetched to disk land in a shared temp cache ($TMPDIR/google-mcp) that is swept of entries older than 24h.

Meet notes

• meet_create_link creates a standalone Meet space via the Meet REST API v2 (spaces.create) — nothing is placed on the calendar. For a titled, scheduled meeting use calendar_create_event (with conferenceData) or calendar_add_meet_link.
• Scope requirement: spaces.create needs the https://www.googleapis.com/auth/meetings.space.created OAuth scope. Existing accounts were authorized before this scope was requested and currently get 403 — Request had insufficient authentication scopes from meet_create_link (verified live). Re-run add-google-account.sh (external openbrain repo) for the account slug to re-consent with the new scope.
• Conference records and transcripts require Google Workspace (transcripts: Business Standard or higher).

Development

npm install
npm run build     # tsc -> dist/
npm test          # vitest run
npm run dev       # tsx src/index.ts (needs --slug)

CI runs tsc --noEmit and the vitest suite on every push/PR.