PRivet (Python Rivet)

PRivet is a fork of Rivet focused on running the full graph executor in Python while keeping the TypeScript UI. The UI still consumes declarative specs; execution now happens in packages/privet via a WebSocket endpoint that mirrors the UI’s flow.

Current Status

• All nodes ported to Python with WS-driven tests (85/85 passing under packages/privet/privet/tests/nodes). Tests load UI-format .rivet-project graphs and talk to the same backend WS endpoint as the UI.
• Chat nodes: real OpenAI-compatible chat calls when OPENAI_API_KEY is set; deterministic fallback (including tool-calls) when no key is present.
• MCP nodes: real MCP HTTP and stdio support. HTTP uses the provided serverUrl; stdio spawns the configured server from project metadata. Synthetic fallback remains for offline runs.
• Subgraph, loop, dataset/vector store, text/object/array transforms, control-flow, and utility nodes all execute in Python with parity-focused behavior.

How to Run

• Backend: cd packages/privet && uvicorn privet.main:app --reload --port 8000 (WS at ws://localhost:21888).
• Tests: cd packages/privet && pytest privet/tests/nodes.
• The UI (TypeScript) connects to the same WS endpoint; configure via Settings → Remote Debugger Server URL.

MCP Configuration (stdio)

For stdio MCP, include server definitions in project metadata (mirrors TS):

metadata:
  mcpServer:
    mcpServers:
      myServer:
        command: "path/to/server"
        args: ["--flag"]
        env: { KEY: "value" }
        cwd: "/path"

Set transportType: stdio and serverId: myServer on MCP nodes. HTTP transport uses serverUrl.

Notes

• Repo is under active refactor; execution logic lives in packages/privet. Avoid editing generated node specs in the UI.
• Default behavior prefers real services (OpenAI/MCP) when credentials/servers exist; fallbacks keep tests runnable offline.