marimo-mcp

marimo-mcp

Auto-discovers running marimo notebooks and exposes tools for reading, editing, and running cells via MCP, supporting both HTTP and VS Code backends.

Category
Visit Server

README

marimo-mcp

A single MCP server that auto-discovers all running marimo notebooks and exposes tools for reading, editing, and running cells — no --mcp flag required.

Works with two backends:

  • HTTP mode — connects to marimo notebooks running via marimo edit (standard server)
  • VS Code mode — connects to marimo notebooks open in VS Code via the companion bridge extension

Architecture

Claude / MCP client
       │
       ▼
  marimo-mcp (Python MCP server)
       │
       ├─── HTTP backend ──────► marimo edit --no-token notebook.py
       │                         (port 2718 by default)
       │
       └─── VS Code backend ───► marimo-mcp-bridge (VS Code extension)
                                  │  port 42018
                                  ▼
                            vscode.commands.executeCommand('marimo.api', ...)
                                  │
                                  ▼
                          marimo VS Code extension

Discovery runs on every tool call (cached 5 s):

  1. Scans running processes for marimo commands, extracts ports
  2. For each port: fetches the HTML page to extract Marimo-Server-Token, then queries /api/home/running_notebooks
  3. Checks if the bridge extension is running on port 42018 and appends any VS Code notebooks

VS Code backend — how cell execution works:

marimo's VS Code extension does not expose an HTTP server. The bridge extension (marimo-mcp-bridge) fills this gap — it runs a small HTTP server inside VS Code and forwards calls through VS Code's notebook execution pipeline (notebook.cell.executeexecuteHandler → marimo LSP → kernel). This means edit_and_run_cell updates the cell visually in VS Code and returns actual stdout/stderr output.

VS Code backend — get_deps and get_variables:

These tools use static analysis of the .py file via marimo's own AST engine. No kernel or bridge needed — just the file on disk. get_variables returns variable names and their kinds (variable, import, function, class); values are only available via the HTTP backend with a running session.


Installation

1. Python MCP server

Requires Python 3.11+ and uv.

git clone <repo>
cd marimo-mcp
uv sync          # creates .venv and installs all dependencies

2. VS Code bridge extension (for VS Code notebooks)

cd marimo-mcp-bridge
npm install
bash install.sh  # compiles TypeScript, packages as VSIX, installs in VS Code

After installing, reload VS Code window (Developer: Reload Window).

For subsequent updates after code changes, just run bash install.sh again.


Configuration

MCP settings

Add to .vscode/mcp.json or Claude Code's MCP config:

{
  "mcpServers": {
    "marimo": {
      "command": "uv",
      "args": ["run", "marimo-mcp"],
      "cwd": "/path/to/marimo-mcp",
      "env": {
        "MARIMO_TOKEN": "optional — only needed if marimo started with token auth"
      }
    }
  }
}

uv run automatically uses the .venv created by uv sync.

Token authentication

By default marimo generates a random access token. Either:

  • Start marimo with --no-token to disable authentication, or
  • Set MARIMO_TOKEN to the token from the startup URL (?access_token=...)

Tools

Tool HTTP VS Code Description
list_notebooks List all discovered notebooks
create_notebook Create a new .py notebook file
get_cells List cells with IDs and code
get_cell_outputs Visual output and console streams
get_errors All errors grouped by cell
get_variables ✓ (with values) ✓ (names/kinds only) Variables in the notebook
get_deps Cell dependency graph
add_cell Add a new cell (not executed); cell_type="markdown" for markup
edit_and_run_cell Edit a cell and run it, returns stdout/stderr
delete_cell Delete a cell

get_cell_outputs and get_errors return an explicit error for VS Code notebooks. Use edit_and_run_cell with print() calls to inspect values.

add_cell parameters

add_cell(notebook, code, after_cell_id=None, cell_type="code")
  • cell_type="code" — standard Python cell (default)
  • cell_type="markdown" — markdown cell; in VS Code uses native NotebookCellKind.Markup (renders immediately without execution); in HTTP mode wraps in mo.md(...)

Current limitations

  • VS Code output is stdout/stderr only — rich outputs (plots, dataframes, marimo UI elements) are not captured via the bridge. Use the HTTP backend (marimo edit --no-token) for full output access.

Claude Code skill

A Claude Code skill for working with this MCP server is available at ~/.claude/plugins/marketplaces/marimo-mcp/SKILL.md. It's enabled automatically when the marimo-mcp@marimo-mcp plugin is active in your Claude Code settings.

This skill is complementary to marimo-pair (which handles marimo edit HTTP mode). Use marimo-mcp when the notebook is open in VS Code.


Testing guide

Test 1: HTTP backend (marimo running locally)

Start a notebook:

marimo edit --no-token --port 2718 /tmp/test_notebook.py

Verify discovery:

uv run python -c "
import asyncio
from marimo_mcp.discovery import discover_notebooks
async def main():
    nbs = await discover_notebooks()
    for nb in nbs:
        print(f'{nb.name}  port={nb.port}  via={\"vscode\" if nb.is_lsp else \"http\"}')
asyncio.run(main())
"

Edit and run a cell:

import asyncio
from marimo_mcp.server import get_cells, edit_and_run_cell

async def main():
    cells = await get_cells('test_notebook.py')
    # get a cell_id from the output
    result = await edit_and_run_cell('test_notebook.py', 'CELL_ID', 'x = 6 * 7\nprint(x)')
    print(result)  # {"output": "42", "stdout": "42\n", ...}

asyncio.run(main())

Test 2: VS Code bridge extension

Verify bridge is running:

curl -s http://127.0.0.1:42018/health
# {"status":"ok"}

List open VS Code notebooks:

curl -s http://127.0.0.1:42018/notebooks | python3 -m json.tool

Full round-trip (edit + run + get output):

import asyncio
from marimo_mcp.server import get_cells, edit_and_run_cell

async def main():
    cells = await get_cells('goyda.py')  # VS Code notebook
    cell_id = ...  # from cells output
    result = await edit_and_run_cell('goyda.py', cell_id, 'print(6 * 7)')
    print(result)  # {"output": "42", "stdout": "42\n", "stderr": ""}

asyncio.run(main())

Test 3: Unit tests

uv run pytest tests/ -v

25 tests should pass, covering MarimoClient, discovery logic, and notebook creation.


Troubleshooting

No notebooks found, but marimo is running:

  • Run with --no-token, or set MARIMO_TOKEN
  • Confirm the port is accessible: curl http://localhost:2718/

Bridge not available (connection refused on port 42018):

  • Check the VS Code Output panel for "marimo-mcp-bridge" channel
  • Make sure a .py marimo notebook is open — the marimo.api command is only available when the marimo extension is active

Bridge needs reinstalling after code changes:

cd marimo-mcp-bridge
bash install.sh
# Then: Developer: Reload Window in VS Code

edit_and_run_cell returns empty output or times out (VS Code):

The cell execution uses VS Code's notebook pipeline. If it times out (15s default):

  1. Check VS Code Output → marimo for kernel startup errors
  2. Make sure the notebook is open and visible (not just in the background)
  3. Try running a cell manually first to warm up the kernel

Wrong Python executable (kernel fails to start):

The bridge resolves Python in this order:

  1. .venv/bin/python next to the notebook file
  2. .venv/bin/python in any VS Code workspace folder
  3. VS Code Python extension active environment
  4. python3 (system fallback)

Create a .venv with marimo in the workspace root:

python3 -m venv .venv
.venv/bin/pip install marimo

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured