A
licenseB
qualityF
maintenanceEnables IDE and DeepSeek Harness agents to control an agent-kernel instance over stdio, exposing tools for auth, projects, assignments, runs, scheduler nudges, and executor settings.
12
MIT
machine-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@machine-mcp Run `git status` in the workspace and summarize the changes."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
machine-mcp
A laptop-side MCP server. An AI agent talks to it over stdio and runs tools on the human's own machine (the one already signed into Copilot CLI and the rest of the local toolchain).
This is a local command process. It does not wrap Cursor or Grok Bot pairing. It does not replace Grok Bot Computers or Local execution.
Install
Python 3.10 or newer.
cd machine-mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -e .Dev tests:
pip install -e ".[dev]"
pytest tests/Related MCP server: local-mcp
Run
machine-mcpstdio servers block on stdin. The client starts the process and owns the pipes.
Smoke test (prints status JSON, exits 0):
machine-mcp --checkOn start, if MACHINE_MCP_TOKEN is unset, a random bearer token is generated and printed once to stderr:
machine-mcp token: ...Keep that token for HTTP v2. stdio v1 does not check it: the client already spawned this process as you.
Cursor / Grok Bot connector
Add a local MCP server whose command is the venv entry point. No pairing protocol.
{
"mcpServers": {
"machine-mcp": {
"command": "path/to/.venv/bin/machine-mcp"
}
}
}See examples/mcp.json. Point command at the absolute path of .venv/bin/machine-mcp on this laptop.
Optional env:
MACHINE_MCP_WORKSPACEfile-tool root. Default:~/machine-mcp-workspace(created on start).MACHINE_MCP_TOKENbearer token for a future HTTP listener. Generated if unset.
Tools
Tool | What it does |
| OS, hostname, cwd, Python version, whether |
|
|
| Honest stub that never writes a video file. On Linux with |
| Write a path relative to the workspace. Refuses |
| Read a path relative to the workspace. Same sandbox. 1MB limit. |
Local only
stdio is local to this process. There is no listen socket in v1. HTTP is optional later and must require MACHINE_MCP_TOKEN. Do not expose this server on a network.
License
MIT. Copyright Andrea Griffiths 2026.
Available Tools
5 toolsread_fileA
Read a UTF-8 file relative to the workspace root. Size limit 1MB.
Refuses absolute paths and parent-directory segments.
Inputs: path (required). Returns: object
record_terminalA
Record a bash script as a terminal demo.
This host does not start a recorder. If ffmpeg is missing, returns an error. If ffmpeg is present, returns the command you would run (Linux x11grab) and does not write a video file.
Inputs: script (required). Returns: object
runA
Run a command in the workspace root (not a shell).
Args: command: Program and arguments as a single string. Split with shlex on POSIX. timeout_seconds: Kill after this many seconds. Default 30, max 120.
Output is limited to 1MB each for stdout and stderr. A nonzero exit code returns ok=false.
Inputs: command (required), timeout_seconds. Returns: object
statusA
Report OS, hostname, cwd, Python version, and whether copilot, ffmpeg, and git are on PATH.
Does not return secrets, tokens, or a home-directory listing.
Inputs: no named parameters. Returns: object
write_fileA
Write a UTF-8 file relative to the workspace root.
Refuses absolute paths and parent-directory segments.
Inputs: path (required), content (required). Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
A4.1/5.0
Scored across 5 tools
Disambiguation4/5
Each tool targets a distinctly different action: environment status, command execution, terminal recording, and file read/write. The only confusing tool is record_terminal, which despite its name does not actually record anything and instead returns a command to run.
Naming Consistency4/5
Most tools follow a clear verb_noun pattern: record_terminal, write_file, read_file, plus the imperative run. status is a minor outlier as a noun, but the overall naming scheme is predictable and readable.
Tool Count4/5
Five tools is a reasonable, well-scoped size for a machine and workspace automation server. However, record_terminal feels loosely connected to the rest of the set, so the count is appropriate but not perfectly cohesive.
Completeness4/5
Core workflows are covered: checking environment state, running commands, and reading/writing files are sufficient for many workspace tasks. Directory listing and deletion are missing as dedicated tools, but agents can work around these gaps using run.
Maintenance
ActivityStale
ResponsivenessNo recent community issues
Related MCP Connectors
Operate Linux, macOS and Windows from your LLM. Every action runs through an auditable allowlist.
Connect ChatGPT, Claude, Cursor or VS Code to your Windows PC: files, terminal, browser, apps.
Connect AI assistants to your approved computer for files, commands, browser and desktop tasks.
Develop, manage, and debug Railway projects, services, and deployments from within agents.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables AI assistants to securely inspect, modify, and manage local projects with trusted filesystem access, safe Git operations, controlled task execution, and developer runtime process management over stdio or HTTP.4Apache 2.0
- AlicenseNot gradedqualityCmaintenanceEnables remote AI agents to read, patch, and (opt-in) execute commands in a local workspace, with all actions approved locally and fully audited.2,672 npm2MIT
- AlicenseNot gradedqualityBmaintenanceEnables local, permissioned development workflows from compatible AI clients, including filesystem editing, interactive terminals, durable jobs, Git operations, and semantic browser testing.64 npm1Apache 2.0