About Enoch
enoch-mcp is a local stdio MCP server for MCP clients that need to inspect and operate a configured Enoch control plane. It forwards typed tool calls to the Enoch API, with authentication and operation metadata rather than raw shell access.
Key features
- 01Typed tools for control-plane and dashboard endpoints
- 02Bearer-token API authentication
- 03Read-only annotations and approval metadata
- 04Optional CPU/GPU worker diagnostics
- 05No raw shell or arbitrary SSH commands
Use cases
- 01Inspecting Enoch control-plane status from an MCP client
- 02Operating Enoch through typed API tools
- 03Checking configured worker health and diagnostics
MCP support
It exposes Enoch control-plane, Dashboard V1, and core API operations as MCP tools over local stdio. MCP clients receive API responses; optional worker probes provide named diagnostics.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
uvx enoch-mcpENOCH_API_URLBase URL for the Enoch API.ENOCH_API_TOKENSecretBearer token for the Enoch API.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEenoch-mcp
enoch-mcp is a local Model Context Protocol stdio server for a configured Enoch FastAPI control-plane API. It lets MCP clients inspect and operate Enoch through typed tools without exposing a raw shell or reimplementing Enoch business logic.
Enoch system repo · Docs · MCP
What it does
- Registers MCP tools for Enoch control-plane, Dashboard V1, and core endpoints.
- Sends requests to a configured Enoch API URL.
- Adds
Authorization: Bearer <token>to API requests using the configured token. - Returns Enoch API responses to the MCP client.
- Marks read-only tools with MCP read-only annotations.
- Marks mutating tools as non-read-only and adds approval metadata.
- Keeps safe defaults for dry-run operations.
- Optionally probes configured CPU/GPU workers directly through worker APIs or allowlisted SSH diagnostics.
What it does not do
- It does not expose a raw shell tool.
- It does not accept arbitrary SSH commands.
- It does not read or write local artifact files directly.
- It does not call language models.
- It does not cache, retry, queue, or schedule work.
- It does not bypass Enoch authentication or authorization.
Requirements
- Python 3.11 or newer
- A running Enoch API, normally at
http://localhost:8787 - An Enoch API bearer token
- An MCP client that can run local stdio servers
Installation
Run from PyPI with uvx:
uvx enoch-mcp --api-url http://localhost:8787 --api-token '<token>'
Or configure with environment variables:
export ENOCH_API_URL='http://localhost:8787'
export ENOCH_API_TOKEN='<token>'
uvx enoch-mcp
For local development from a checkout:
git clone https://github.com/alias8818/enoch-mcp.git
cd enoch-mcp
uv sync --dev
uv run enoch-mcp --api-url http://localhost:8787 --api-token '<token>'
Configuration
| Option | Environment variable | Default | Description |
|---|---|---|---|
--api-url | ENOCH_API_URL | http://localhost:8787 | Base URL for the Enoch API. |
--api-token | ENOCH_API_TOKEN | none | Bearer token for the Enoch API. |
--worker-probes-json | ENOCH_WORKER_PROBES_JSON | none | Optional JSON map for direct worker diagnostics. |
--worker-probes-file | ENOCH_WORKER_PROBES_FILE | none | Optional path to a JSON map for direct worker diagnostics. |
The token is required. If it is missing, tool calls fail before making an HTTP request.
Optional worker probes
Worker probes are disabled unless ENOCH_WORKER_PROBES_JSON or ENOCH_WORKER_PROBES_FILE is configured. This keeps the default package a thin control-plane bridge. When configured, the MCP exposes named diagnostics for worker truth: API health, worker-gate dashboard status, active process markers, bounded log tails, disk space, and expected artifact presence.
Example:
{
"cpu": {
"api_url": "http://127.0.0.1:18788",
"api_token": "worker-api-token",
"service_name": "enoch-control-plane",
"project_root": "/srv/enoch/projects"
},
"gpu": {
"api_url": "http://127.0.0.1:18789",
"api_token": "worker-api-token",
"ssh_host": "worker-gpu.example.internal",
"ssh_user": "enoch",
"service_name": "enoch-control-plane",
"project_root": "/srv/enoch/projects",
"log_paths": ["/var/log/enoch-control-plane.log"]
}
}
Supported fields per lane:
api_url: worker-gate base URL. Used first for/healthz,/dashboard/api,/dashboard/api/run/{run_id}, and/project-status/{project_id}.api_token: worker bearer token. Treated as secret.ssh_host,ssh_user,ssh_port: optional SSH fallback/debug target.service_name: systemd unit name for service checks and journal tails.project_root,state_dir: fixed worker roots used for disk and artifact checks.log_paths: fixed worker-gate log paths that may be tailed.
SSH probes only run fixed diagnostic commands. They do not accept arbitrary shell input from the MCP client. User-supplied IDs are limited to safe run/project identifier characters, log output is bounded, SSH uses batch mode and no stdin, and the recommended deployment is a read-only worker user or forced-command policy.
MCP client setup
Claude Desktop
{
"mcpServers": {
"enoch": {
"command": "uvx",
"args": ["enoch-mcp"],
"env": {
"ENOCH_API_URL": "http://localhost:8787",
"ENOCH_API_TOKEN": "replace-with-token"
}
}
}
}
For local development, point the MCP client at the checkout:
{
"mcpServers": {
"enoch": {
"command": "uv",
"args": ["--directory", "/path/to/enoch-mcp", "run", "enoch-mcp"],
"env": {
"ENOCH_API_URL": "http://localhost:8787",
"ENOCH_API_TOKEN": "replace-with-token"
}
}
}
}
Use the equivalent local stdio-server settings for Cursor, Copilot, Windsurf, or other MCP clients.
Development
uv sync --dev
uv run pytest -q
gitleaks detect --no-git --redact
Public-safety note
This package is public. Examples must use placeholders and local URLs only. Do not publish private hostnames, LAN/Tailscale IPs, operator paths, or live tokens.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.