mcp-monitor-local-sdk

# mcp-monitor-local-sdk Self-host monitoring SDK for Python MCP servers. [User guide](#user-guide) · [Developer guide](#developer-guide) Repo copies: [USER.md](USER.md) · [DEV.md](DEV.md) ## User guide Point this SDK at your self-host API (`metrics_server_url` + `instance_secret`). It wraps tool handlers and sends tool-call metrics (name, duration, success/error) — not application logs — to `POST /v1/metrics`. Open the local UI or desktop app to see them. This package does not talk to the hosted dashboard. ```bash pip install mcp-monitor-local-sdk ``` ```python from mcp_monitor_sdk import MonitoredFastMCP server = MonitoredFastMCP( "todo-mcp", server_name="todo-mcp", instance_secret="change-me-to-a-long-random-secret", metrics_server_url="http://localhost:8000/v1/metrics", ) @server.tool() def todos_list() -> str: """List all todos.""" return "No todos." if __name__ == "__main__": server.run() ``` `instance_secret` must match `INSTANCE_SECRET` on the API. ## Developer guide Same ingest contract as the JS SDK. Built to sit next to the official [`mcp`](https://pypi.org/project/mcp/) package. ### Options | Option | Required | Default | Notes | |---|---|---|---| | `server_name` | yes | | Trimmed and lowercased | | `instance_secret` | yes | | Min 16 characters; sent as `X-Instance-Secret` | | `metrics_server_url` | yes | | Full ingest URL, usually `http://localhost:8000/v1/metrics` | | `batch_size` | no | 10 | Max 100 | | `retry_attempts` | no | 2 | `0` still tries once | | `timeout_ms` | no | 5000 | | | `flush_interval_ms` | no | 5000 | | | `log_level` | no | `info` | `debug` \| `info` \| `warn` \| `error` \| `silent` | Events flush when the batch is full or every `flush_interval_ms`. Failed flushes are restored. Shutdown drains remaining events. Local checkout: ```bash pip install -e ./local/sdk/py ``` Publish: ```bash cd local/sdk/py python -m build twine upload dist/* ```