Mcp Server Colab Exec

io.github.pdwi2020Published by GitHub account pdwi2020v0.1.0Updated Oct 24, 2026

Execute Python code on Google Colab GPU runtimes (T4/L4) from any MCP client

VerifiedSTDIODesktop onlyCloud & InfrastructureAI & ML

Overview

AI-generated overview

A local MCP server that allocates Google Colab GPU runtimes and runs Python code on them, giving an assistant GPU compute without local hardware.

What you can do
Run inline Python code on a Colab GPU runtime, choosing a T4 or L4 accelerator and a timeout, and get back per-cell output, errors and stderr as JSON. Run a local .py file from the current workspace on a Colab GPU runtime. Execute code and download generated artifacts such as images, CSVs or model weights into a local output directory.
When to use it
Useful when an assistant needs to run GPU-accelerated Python (CUDA, PyTorch, TensorFlow) or train and export models, but the machine has no GPU. Also handy for quick checks of GPU availability or nvidia-smi output on a Colab runtime.
Requirements
Runs locally as a Python package (Python 3.10+) via pip or uvx, configured as a stdio server in the MCP client. Needs a Google account with Colab access and a browser for the first-run Google OAuth2 consent; the token is cached locally and refreshed automatically. Network access to Colab is required.
Before you install
Code is executed on a remote Google Colab runtime, so anything in the code or file is sent to Google. The OAuth token and refresh token are cached on disk; treat that file as a credential. L4 accelerators are described as premium, and Colab enforces usage quotas. Artifact downloads write files into a local output directory.

Installation

In SourceWeft

  1. Open Mcp Server Colab Exec in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

mcp-server-colab-exec

MCP server that allocates Google Colab GPU runtimes (T4/L4) and executes Python code on them. Lets any MCP-compatible AI assistant — Claude Code, Claude Desktop, Gemini CLI, Cline, and others — run GPU-accelerated code (CUDA, PyTorch, TensorFlow) without local GPU hardware.

Prerequisites

  • Python 3.10+
  • A Google account with access to Google Colab
  • On first run, a browser window opens for OAuth2 consent. The token is cached at ~/.config/colab-exec/token.json for subsequent runs.

Installation

bash
pip install mcp-server-colab-exec

Or run directly with uvx:

bash
uvx mcp-server-colab-exec

Configuration

Claude Code

Add to your project's .mcp.json or ~/.claude/.mcp.json:

json
{  "mcpServers": {    "colab-exec": {      "command": "mcp-server-colab-exec"    }  }}

Or via the CLI:

bash
claude mcp add colab-exec mcp-server-colab-exec

Claude Desktop

Add to claude_desktop_config.json:

json
{  "mcpServers": {    "colab-exec": {      "command": "mcp-server-colab-exec"    }  }}

Gemini CLI

bash
gemini mcp add colab-exec -- mcp-server-colab-exec

Tools

colab_execute

Execute inline Python code on a Colab GPU runtime.

ParameterTypeDefaultDescription
codestring—Python code to execute (required)
acceleratorstring"T4"GPU type: "T4" (free) or "L4" (premium)
timeoutint300Max execution time in seconds

Returns JSON with per-cell output, errors, and stderr.

colab_execute_file

Execute a local .py file on a Colab GPU runtime.

ParameterTypeDefaultDescription
file_pathstring—Path to a local .py file (required)
acceleratorstring"T4"GPU type: "T4" (free) or "L4" (premium)
timeoutint300Max execution time in seconds

Security policy: file_path must be a .py file inside the current workspace (cwd).

colab_execute_notebook

Execute code and collect all generated artifacts (images, CSVs, models, etc.).

ParameterTypeDefaultDescription
codestring—Python code to execute (required)
output_dirstring—Local directory for downloaded artifacts (required)
acceleratorstring"T4"GPU type: "T4" (free) or "L4" (premium)
timeoutint300Max execution time in seconds

Artifacts are downloaded as a zip and extracted into output_dir. Zip members are validated before extraction to prevent path traversal and special-file writes.

Examples

Check GPU availability:

colab_execute(code="import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))")

Run nvidia-smi:

colab_execute(code="import subprocess; print(subprocess.run(['nvidia-smi'], capture_output=True, text=True).stdout)")

Train a model and download weights:

colab_execute_notebook(    code="import torch; model = torch.nn.Linear(10, 1); torch.save(model.state_dict(), '/tmp/model.pt')",    output_dir="./outputs")

Authentication

On first use, the server opens a browser window for Google OAuth2 consent. The access token and refresh token are cached at ~/.config/colab-exec/token.json. Subsequent runs use the cached token and refresh it automatically.

The OAuth2 client credentials are the same ones used by the official Google Colab VS Code extension ([email protected]). They are intentionally public.

Troubleshooting

"GPU quota exceeded" — Colab has usage limits. Wait and retry, or use a different Google account.

"Timed out creating kernel session" — The runtime took too long to start. Retry — Colab sometimes has delays during peak usage.

"Authentication failed" — Delete ~/.config/colab-exec/token.json and re-authenticate.

OAuth browser window doesn't open — Ensure you're running in an environment with a browser. For headless servers, authenticate on a machine with a browser first and copy the token file.

License

MIT

Source: README.md at commit b67d0d7

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.0LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts