python-repl

Python REPL MCP 3.x

Load execute_python through tool search when deferred. Use python_health to verify server version, selected Python interpreter, limits and broker configuration. For a long cell use python_start to obtain a run ID immediately, then poll or cancel. For MCP 2026-07-28, take session_id from health and pass it to execution; legacy clients can omit it. One cell runs at a time per session. python_run polls progress/result and python_cancel cancels an active run independently of Python execution.

For parallel agents or a project-specific Python environment, open a dedicated session with python_session_open(name, project, python?) and pass its session_id to execution, polling and cancellation. Different sessions have independent variables, cwd, active runs and history. Names are labels, not identifiers: retain the returned ID. Inspect the reported environment; explicit Python takes precedence, an existing project .venv is used otherwise, and projects without .venv use the reported server default. Invalid explicit/project environments fail instead of falling back.

Use python_session_inspect(session_id, include_namespace=True) after reconnecting or losing conversation context. It returns bounded names/types without custom repr. Pass expected_generation from the inspected state to execution when code depends on existing helpers. STATE_CHANGED means the code was not executed: inspect state and recreate needed helpers explicitly. Never replay code automatically after an unknown write outcome. Close only your own session with python_session_close.

Use python_execute_file(session_id, path, argv=[], mode="persistent") for a saved Python script in the selected environment. mode="fresh" executes in a separate owned worker and preserves the persistent namespace. Both modes retain output, run identity and cancellation semantics. Relative file paths resolve against the session project; source tracebacks identify the actual file.