# Turnstone MCP Server

A [Model Context Protocol](https://modelcontextprotocol.io/) server providing five sandboxed tools for LLM-driven workflows. Built with **FastMCP** and **Playwright**, it exposes code execution, headless browser automation, file-to-Markdown conversion, and ephemeral file downloads over HTTP.

Designed to run as a Docker container alongside the Turnstone platform, sharing the `turnstone_workspace` volume so tools can read and write files that other services in the ecosystem use.

## Tools

| Tool | Description |
|---|---|
| `execute_python_code` | Runs Python source in a subprocess with timeout, memory, and output-size limits |
| `playwright_navigate` | Navigates a persistent headless Chromium session to a URL |
| `playwright_get_page_text` | Extracts rendered text from the current browser page |
| `extract_to_markdown` | Converts supported files (PDF, DOCX, images, CSV, etc.) to Markdown via `markitdown` |
| `generate_download_link` | Creates a single-use, time-limited HTTPS download URL for a file in `/workspace` |

## Quick Start

Run the server with Docker Compose:

```bash
docker compose up --build
```

The server starts on port **8000**. The MCP health endpoint is available at `http://localhost:8000/health`.

### Running in the background

```bash
docker compose up -d --build
```

## Configuration

All limits are configurable via environment variables with sensible defaults. Set them in your `.env` file or pass them through `docker compose`:

| Variable | Default | Description |
|---|---|---|
| `PYTHON_EXEC_TIMEOUT` | `30` | Python execution timeout (seconds) |
| `PYTHON_MAX_OUTPUT` | `65536` | Maximum stdout/stderr capture (bytes) |
| `PYTHON_MAX_MEMORY_MB` | `256` | Address-space memory limit per subprocess (MB) |
| `PLAYWRIGHT_NAV_TIMEOUT` | `30000` | Browser navigation timeout (ms) |
| `PLAYWRIGHT_TEXT_TIMEOUT` | `15000` | Page text extraction timeout (ms) |
| `EXTRACT_MAX_FILE_SIZE` | `10485760` | Maximum file size for extraction (10 MB) |
| `PUBLIC_BASE_URL` | `http://localhost:8000` | Public HTTPS base URL for ephemeral download links |
| `WORKSPACE_MOUNT` | `/workspace` | Absolute path to the shared workspace root (used as the Docker volume name in `compose.yaml`) |

Example with custom limits:

```bash
PYTHON_EXEC_TIMEOUT=60 PYTHON_MAX_MEMORY_MB=512 docker compose up --build
```

## Architecture

```
┌──────────────┐     HTTP/MCP      ┌──────────────────┐
│   LLM Client │ ────────────────► │  Turnstone MCP    │
│              │                   │  Container :8000  │
└──────────────┘                   └────────┬─────────┘
                                            │
                    ┌───────────────────────┼───────────────────────┐
                    │                       │                       │
              ┌─────▼─────┐          ┌──────▼──────┐       ┌───────▼───────┐
              │ subprocess │          │ Playwright  │       │ markitdown    │
              │ (Python)   │          │ Chromium    │       │ (file → MD)   │
              └───────────┘          └─────────────┘       └───────────────┘
                    │                       │                       │
                    └───────────┬───────────┴───────────────────────┘
                                │
                      ┌─────────▼──────────┐
                      │  turnstone_workspace│
                      │  (shared volume)    │
                      └────────────────────┘
```

### Key design decisions

- **Persistent browser session** — A singleton `PlaywrightBrowserManager` holds the Chromium process alive across tool calls, preventing garbage-collection teardowns.
- **Sandboxed execution** — Python code runs in a separate process with a per-subprocess memory limit (`resource.setrlimit`) and a wall-clock timeout enforced by `subprocess.run`.
- **Path traversal protection** — `extract_to_markdown` and `generate_download_link` resolve all paths against `WORKSPACE_MOUNT` and reject anything that escapes it.
- **Ephemeral download links** — `generate_download_link` mints a 256-bit single-use token linked to a file path; clicking the URL streams the file and immediately invalidates the token. Expired tokens are garbage-collected automatically.
- **Structured JSON logging** — All tool invocations and responses are logged as structured JSON for observability.
- **CORS enabled** — The server accepts requests from any origin with standard MCP headers.
- **Shared workspace volume** — The container mounts the `turnstone_workspace` Docker volume at `/workspace`, giving the MCP tools read/write access to files used by other Turnstone services.

## Credential Leak Guardrail — Action Required

The Turnstone credential leak guardrail intercepts outbound messages and flags URLs it considers suspicious. Because `generate_download_link` returns a URL containing a token, the guardrail will block it unless a policy explicitly allows the pattern.

> **Action:** Add the following pattern to the Turnstone credential leak policy:
>
> ```
> PUBLIC_BASE_URL/download?
> ```
>
> Replace `PUBLIC_BASE_URL` with the actual value of the environment variable (e.g., `https://turnstone.example.com/download?`).

Without this policy, any attempt by the agent to present a download link to the user will be silently blocked by the credential leak filter.

## Docker Configuration

The service is defined in `compose.yaml`:

| Setting | Value |
|---|---|
| Port mapping | `8000:8000` |
| User | `1000:1000` |
| Restart policy | `unless-stopped` |
| Health check | HTTP GET `/health` every 30 s (timeout 10 s, 3 retries) |
| Workspace mount | `${WORKSPACE_MOUNT:-workspace}:/workspace` |

## Endpoints

| Method | Path | Description |
|---|---|---|
| `POST` | `/` | MCP protocol endpoint (default) |
| `GET` | `/health` | Returns `{"status": "healthy", "service": "turnstone-mcp"}` |
| `GET` | `/download?token=` | Streams a file identified by a single-use token (see `generate_download_link` tool) |

## Project Structure

```
compose.yaml       # Docker Compose service definition
Dockerfile         # Container build instructions
server.py          # Complete MCP server — all logic in a single module
```

## Development

No test framework or build step is required. To add a new tool, follow the existing pattern:

1. Define the function with `@mcp.tool()`
2. Add docstring with `Parameters` and `Returns` sections
3. Implement input validation and error handling
4. Return a string result (success or error message)

## License

See the repository root for licensing information.