Skip to main content
Pre-release

This release is a pre-release and may not be stable for production use.

mcp-gee-sweet logo

mcp-gee-sweet

A Google Workspace MCP server built for depth.

GitHub License

mcp-gee-sweet gives AI clients deep, direct access to Google Sheets, Docs, Drive, Calendar, and Gmail. It reads raw formulas like =VLOOKUP(), builds Docs structurally (colspan/rowspan tables, themes, named styles, an HTML/Markdown→AST→Docs pipeline), and covers Drive files, sharing, and sync alongside Calendar events and Gmail messages, threads, and drafts.

A few highlights:

  • Caching — repeated reads return in ~2ms from a local cache instead of a round-trip API call every time.
  • Formula access — reads raw =VLOOKUP(), =SUM(), and =IF() expressions, not just the values they produce.
  • Local sync — two-way sync between a local folder and Google Drive (add the local-files tool group to ENABLED_TOOLS; see Choosing which tools load).
A Markdown file in an editor beside the Google Doc that create_doc_from_file produces from it, with the heading, table, nested list, quote, and code block intact

Install (stable):

uvx mcp-gee-sweet

Install (bleeding edge — every code change on develop):

uvx --prerelease=allow mcp-gee-sweet

Full documentation →


What's covered

Service Tools Highlights
Google Sheets 36 Formula access, batch read/write, 8 chart types, structural ops
Google Drive 38 File ops, sharing/permissions, local sync, revision history
Google Docs 34 Full HTML→Doc pipeline, tables, range styling, themes, Markdown export
Google Calendar 17 Event CRUD, free-slot finder across multiple calendars
Gmail 15 Search and read messages and threads, send and reply, drafts you can review, edit, or discard before sending, labels (OAuth only). By default nothing is sent until the user confirms it; a client that can't ask leaves the email as a draft. See Gmail sending

Quick start

Option A: uvx in Claude Desktop — the fastest way in

This path goes from nothing to a first tool call in Claude Desktop on macOS. It runs the published package with uvx, so there's nothing to clone.

1. Install uv. It provides the uvx command. In a terminal:

curl -LsSf https://astral.sh/uv/install.sh | sh

(or brew install uv; other methods are in uv's installation guide). Open a new terminal afterward so uvx is on your PATH, then check it with uvx --version.

2. Create a Google OAuth client. Follow Method B in Authentication, steps 1–4: they create a Google Cloud project, turn on the APIs, set up the consent screen, and download your OAuth client file. Then move the file to a folder of its own:

mkdir -p ~/mcp-gee-sweet
mv ~/Downloads/client_secret_*.json ~/mcp-gee-sweet/credentials.json

3. Sign in to Google once. This opens your browser to Google's consent page and saves a token next to the client file:

CREDENTIALS_PATH=$HOME/mcp-gee-sweet/credentials.json \
TOKEN_PATH=$HOME/mcp-gee-sweet/token.json \
uvx mcp-gee-sweet auth

For a develop build, use uvx --prerelease=allow mcp-gee-sweet auth instead. If Google says it hasn't verified the app, see step 6 of Method B.

4. Find uvx's full path. Claude Desktop doesn't use your shell's PATH, so its config needs the full path:

command -v uvx    # e.g. /Users/you/.local/bin/uvx

5. Add the server to Claude Desktop. In Claude Desktop, open Settings → Developer → Edit Config. That shows you the config file, ~/Library/Application Support/Claude/claude_desktop_config.json; open it in a text editor. Add this, using the path from step 4 and your own home folder in place of /Users/you:

{
  "mcpServers": {
    "mcp-gee-sweet": {
      "command": "/Users/you/.local/bin/uvx",
      "args": ["mcp-gee-sweet"],
      "env": {
        "CREDENTIALS_PATH": "/Users/you/mcp-gee-sweet/credentials.json",
        "TOKEN_PATH": "/Users/you/mcp-gee-sweet/token.json",
        "ENABLED_TOOLS": "default,local-files"
      }
    }
  }
}

If the file already has an mcpServers block, add the "mcp-gee-sweet" entry inside it instead of adding a second block. Write full paths: the server doesn't expand ~. For a develop build, use "args": ["--prerelease=allow", "mcp-gee-sweet"]. ENABLED_TOOLS picks the tools: default,local-files is the everyday set plus the tools that work on files on your Mac, such as create_doc_from_file and sync_folder (see Choosing which tools load).

6. Restart Claude Desktop and try it. Quit Claude Desktop completely (⌘Q, not just closing the window), reopen it, and ask something like "List my most recently modified Google Drive files."

Without step 3's token, the server still starts, and its first tool call returns a Google consent link instead: open it, approve, and ask again.

Running uvx mcp-gee-sweet by itself in a terminal starts a stdio server. It logs a version line (and, with no token, a warning) and then waits silently for an MCP client to talk to it over stdin. Nothing else will happen, so press Ctrl+C: the server is meant to be started by your MCP client, as in step 5. With no credentials at all, it exits instead, with All authentication methods failed. Please configure credentials.

On Windows, the config file is %APPDATA%\Claude\claude_desktop_config.json and where uvx gives uvx's path, but Windows isn't tested yet.

Option B: uv with a cloned repo — local development

To work on the server's code or run your own changes, clone the repo. This needs uv too (step 1 above). No Docker required.

git clone https://github.com/khuisman/mcp-gee-sweet.git
cd mcp-gee-sweet
uv sync

Then point your MCP client at it using stdio transport (see MCP client config below). The client spawns the server as a subprocess on demand — each session gets its own isolated process, so restarting the server for a code change or config update doesn't affect other open sessions.

Option C: Docker — persistent shared server (SSE)

Use Docker when you want a single long-running server that multiple clients connect to over SSE — for example, Claude Desktop talking to the same instance as Claude Code.

git clone https://github.com/khuisman/mcp-gee-sweet.git
cd mcp-gee-sweet
echo "MCP_SERVER_TOKEN=$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')" >> .env
make build   # build the container image
make start   # start the server (SSE on port 47000)
make logs    # tail logs

Point your MCP client at http://localhost:47000/sse, sending the token from .env as an Authorization: Bearer <token> header (see the SSE config below). The server rejects any request without it. Without Docker, --transport sse and --transport streamable-http listen on 127.0.0.1 only, and need MCP_SERVER_TOKEN to serve on any other address; see HTTP transport authentication.

Note: when using SSE, all clients share one server process. After a code change or restart (make restart), you must also restart each MCP client to reconnect.


Configuration

The server tries auth methods in a waterfall by default, OAuth first — it authenticates as you and has full personal Drive access, which is what local use (Options A and B above) needs. For OAuth, download an OAuth Client ID JSON from Google Cloud (how) and point the server at it. Then sign in once, with the command that matches how you run the server:

export CREDENTIALS_PATH="/path/to/credentials.json"
export TOKEN_PATH="/path/to/token.json"

uvx mcp-gee-sweet auth                      # PyPI install, stable
uvx --prerelease=allow mcp-gee-sweet auth   # PyPI install, develop builds
uv run mcp-gee-sweet auth                   # cloned repo, run from its folder

uvx doesn't put an mcp-gee-sweet command on your PATH, so a bare mcp-gee-sweet auth works only if you installed the package some other way (for example uv tool install mcp-gee-sweet).

Pass the same CREDENTIALS_PATH and TOKEN_PATH in your MCP client's env. A stdio server with no token doesn't open a browser itself. Instead its first tool call offers you the Google consent link (or you run the auth command above), and you retry the call. A token revoked mid-session is handled the same way, with no restart.

Service accounts (recommended for headless server deployments, see Option C above), base64 credential injection, and Application Default Credentials are also supported. Gmail tools need OAuth. See Authentication for all options.


MCP client config

Claude Desktop — one-click bundle (new in v0.9.6; on 0.9.5, see its docs): download mcp-gee-sweet-<version>.mcpb from the latest GitHub Release, then double-click it, or install it from Claude Desktop's Settings → Extensions. It needs uv installed first (step 1 of Quick start's Option A). It asks for your OAuth client file and needs no config editing (details).

Claude Desktop — stdio (PyPI install): see step 5 of Quick start's Option A.

Claude Desktop — stdio (cloned repo):

{
  "mcpServers": {
    "mcp-gee-sweet": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mcp-gee-sweet", "mcp-gee-sweet"],
      "env": {
        "CREDENTIALS_PATH": "/path/to/credentials.json",
        "TOKEN_PATH": "/path/to/token.json"
      }
    }
  }
}

Claude Code — stdio (cloned repo):

claude mcp add mcp-gee-sweet \
  -e ENABLED_TOOLS=all \
  -e CREDENTIALS_PATH=/path/to/credentials.json \
  -e TOKEN_PATH=/path/to/token.json \
  -- uv run --directory /path/to/mcp-gee-sweet mcp-gee-sweet

ENABLED_TOOLS=all loads every tool, which suits Claude Code: its tool search loads a tool's full definition only when the model picks it. For a PyPI install, replace everything after -- with uvx mcp-gee-sweet. Run the auth command from Configuration above with the same ENABLED_TOOLS, since it decides the scopes.

Claude Desktop — SSE (Docker):

{
  "mcpServers": {
    "mcp-gee-sweet": {
      "transport": "sse",
      "url": "http://localhost:47000/sse",
      "headers": {
        "Authorization": "Bearer <token>"
      }
    }
  }
}

See Client Setup for more options including tool filtering.

Choosing which tools load

With ENABLED_TOOLS unset, the server loads the default tool group: everyday reading and writing across Sheets, Docs, Drive, Calendar and Gmail, sized for clients that cap how many tools they accept. ENABLED_TOOLS takes tool names and group names, and loads their union:

  • all: every tool. Use it in Claude Code, whose tool search keeps the full set cheap.
  • default,local-files: the default set plus the tools that work on local files: create_doc_from_file and update_doc_from_file for Markdown files, and local upload, download and sync_folder. default leaves these out.
  • A service group (sheets, drive, docs, calendar, gmail) for every tool of that service, e.g. default,docs.

The groups, their sizes and the per-client settings are in Tool filtering.

Using a client with a tool limit

Some MCP clients cap how many tools they accept. The default group is under each current cap below (older Cursor builds capped a session at 40, which default doesn't exceed), but with every tool registered (all) this server is over all of them:

Client Limit
VS Code + GitHub Copilot (agent mode) 128 tools per request, built-in tools included. In our November 2026 test, Copilot looked tools up on demand and handled the full set; a build that sends every tool fails with "Cannot have more than 128 tools per request"
Windsurf 100 tools in total, across all servers
Cursor Depends on the model ("Too many MCP tools are enabled for this mode")
Google Antigravity 100 tools per server (user report)

On Windsurf, Cursor and Antigravity, or on VS Code if you see that error, leave ENABLED_TOOLS unset or use tool groups and names whose total stays under your client's cap; every group but all is under 100 tools. Clients with a tool limit has a ready-to-paste config for VS Code, Cursor, and Windsurf.

Changed in this release: an unset ENABLED_TOOLS used to load every tool, and now loads only the default group. To keep every tool, set ENABLED_TOOLS=all. To keep the local-file tools, add local-files.


Upgrading

Check the changelog before upgrading. A renamed or removed tool or parameter keeps working for one release first, and is listed under that release's Deprecated heading with its replacement. A renamed tool's old name keeps working only when you list it in ENABLED_TOOLS or set ENABLED_TOOLS=all. Update allowlists, ENABLED_TOOLS and saved prompts during that release. The deprecation policy has the details.


Docs


Contributing

See CONTRIBUTING.md for setup, QA workflows, and PR guidelines. A signed-off docs/qa/runs/vX.Y.Z.md record is required before any stable PyPI release — see docs/qa/README.md for the QA workflow.

License

MIT — see LICENSE.

Credits

Metadata

Release files for mcp-gee-sweet 0.9.7.dev50

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for mcp-gee-sweet 0.9.7.dev50
File Size Uploaded Provenance Actions
mcp_gee_sweet-0.9.7.dev50.tar.gz 4.3 MB With provenance Details for mcp_gee_sweet-0.9.7.dev50.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for mcp-gee-sweet 0.9.7.dev50
File Interpreter ABI Platform Actions
mcp_gee_sweet-0.9.7.dev50-py3-none-any.whl Python 3 none any Details for mcp_gee_sweet-0.9.7.dev50-py3-none-any.whl

Total release size: 4.8 MB

Release files / mcp_gee_sweet-0.9.7.dev50.tar.gz

Download URL mcp_gee_sweet-0.9.7.dev50.tar.gz
Size 4.3 MB
Tags Source
SHA-256 checksum
How to use checksums
bb48a92db3d10323cf10784ced9dc491293392989b4f63ac547a935571c9dee5
BLAKE2b-256 checksum
How to use checksums
66311532ae05921a61d1b72eeeec886df0fc7a109b350de0db9cc17c9d15d800
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

Release files / mcp_gee_sweet-0.9.7.dev50-py3-none-any.whl

Download URL mcp_gee_sweet-0.9.7.dev50-py3-none-any.whl
Size 470.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
ef4355fa1b15a91a78044674672adcd3d3065f3a6ea1659f341672a7bc31c985
BLAKE2b-256 checksum
How to use checksums
0c1a4bb63a61b42dae40f6e48a82e6647818f7bf68b0b6a06c2cb954acb72ad0
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

Security for mcp-gee-sweet 0.9.7.dev50

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for mcp_gee_sweet-0.9.7.dev50.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release
Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page