Deploy¶

The server is a pure-Python FastAPI app — no compiler, no Node build. The Studio assets ship inside the wheel.

Docker¶

A Dockerfile is included (installs .[server], runs riptide serve on 0.0.0.0:8000):

docker build -t riptide-watergraph .
docker run -p 8000:8000 riptide-watergraph        # http://localhost:8000/

For a real model, pass a key and install the extra (or extend the image):

docker run -e OPENAI_API_KEY=sk-... -e RIPTIDE_WATERGRAPH_MODEL=gpt-4o-mini \
  -p 8000:8000 riptide-watergraph

Docker Compose¶

A docker-compose.yml mounts a named volume for DATA_DIR (so memory + the usage log persist) and reads env from your shell or a .env file:

docker compose up --build        # http://localhost:8000/

services:
  studio:
    build: .
    ports: ["8000:8000"]
    environment:
      RIPTIDE_WATERGRAPH_DISABLE_TRACING: "1"
      DATA_DIR: "/data"
      # OPENAI_API_KEY / RIPTIDE_WATERGRAPH_MODEL for a real model
    volumes:
      - riptide-data:/data
volumes:
  riptide-data:

Other hosts¶

Render / Railway / Fly.io — deploy the Docker image; expose port 8000; set env vars in the dashboard; attach a persistent volume at your DATA_DIR if you want memory to survive restarts.
Any container platform — the image is a standard python:3.13-slim + riptide serve.

Bind the public listener behind an auth proxy — the API is unauthenticated.

Configuration reference¶

All settings are environment variables (pydantic-settings; a .env file also works).

Variable	Default	Purpose
`RIPTIDE_WATERGRAPH_MODEL`	`gpt-4o-mini`	default LiteLLM model string
`PLANNER_MODEL` / `WORKER_MODEL`	(model)	optional per-role model routing
`DATA_DIR`	`.riptide_watergraph`	base dir for per-tenant memory + usage log
`CHECKPOINT_PATH`	`.riptide_watergraph/checkpoints.sqlite`	LangGraph SqliteSaver DB
`MEMORY_PATH`	`.riptide_watergraph/memory.json`	persistent lesson store
`WORKSPACE_DIR`	`.riptide_watergraph/workspace`	sandbox for the dev/file tools
`TENANT_ID`	`default`	active tenant
`TENANT_BUDGET_USD`	`0` (unlimited)	per-tenant spend ceiling (→ HTTP 402)
`RIPTIDE_MCP_SERVERS`	`[]`	JSON array — the MCP-connect allowlist
`RIPTIDE_WATERGRAPH_DISABLE_TRACING`	`false`	turn off OTEL/Langfuse spans
`LANGFUSE_PUBLIC_KEY` / `LANGFUSE_SECRET_KEY` / `LANGFUSE_HOST`	—	tracing backend

Gated feature flags (off by default): RIPTIDE_ENABLE_ENTERPRISE, RIPTIDE_ENABLE_EXEC, RIPTIDE_ENABLE_NETWORK, RIPTIDE_ENABLE_MCP_CONNECT.

API keys¶

Keys (OPENAI_API_KEY, ANTHROPIC_API_KEY, …) are read by LiteLLM from the environment per call, or set at runtime via the Studio Connections view (in-memory, masked, never written to disk).