PrefectMCP Server

Prefect's MCP server is deliberately narrow: it is read-only. An agent can inspect flow runs, deployments, work pools and automations on Prefect Cloud or a self-hosted instance, but it cannot trigger, pause or delete anything. For a workflow orchestrator that is the right default — the failure mode you want to avoid is an agent "fixing" a stuck run by re-triggering a pipeline that writes to production.

What it is genuinely good at is debugging. The agent reads the run graph, the task states, the logs and the automation that fired, and reasons about why a deployment failed with the same information a human would open the UI to get. Alongside that sits a documentation proxy, so the assistant looks up current Prefect CLI, SDK and deployment guidance and release notes rather than reciting an API shape from two versions ago.

Setup depends on where Prefect lives. Claude Code and Codex both have a plugin (/plugin marketplace add prefecthq/prefect-mcp-server, then /plugin install prefect) that connects to Prefect's hosted read-only server over Prefect Cloud OAuth, with workspace selection during install; because the bundle points at a hosted URL rather than launching a local process, the same plugin works for Claude Tag and Cowork. A local stdio server runs via uvx prefect-mcp and inherits credentials from the active profile in ~/.prefect/profiles.toml, which is the path for self-hosted Prefect or an explicit workspace. Teams can also deploy it over HTTP with per-request credential headers for a shared server.

It is MIT licensed and the maintainers label it experimental and under active development — APIs and behaviours can change without notice, so pin the version if you depend on it.