ApiPi

Documentation: geki-ai.github.io/apipi

ApiPi is an open-source agent platform. Pi runs the agent loop. You host the gateway, keep control of the agents, and point them at your own LLM endpoints. Clients use an HTTP API compatible with the OpenAI Agents API.

Production sessions run in Firecracker microVMs: each guest has its own kernel. The gateway stays on the host. Pi and the session files share the guest.

GEKI also runs a managed ApiPi on European infrastructure.

Quickstart

Python 3.13 and uv. Postgres for the store. The Pi CLI (pi --mode rpc) on PATH for live turns.

pip install geki-apipi
# or: uv add geki-apipi
# S3 artifacts: pip install "geki-apipi[s3]"
npm i -g --ignore-scripts @earendil-works/pi-coding-agent@0.85.1
docker compose up -d postgres
export DATABASE_URL=postgresql+asyncpg://apipi:apipi@localhost:5432/apipi
export OPENAI_BASE_URL=http://your-model-host/v1
apipi migrate
apipi serve

From a checkout, uv sync then prefix commands with uv run. That binds 0.0.0.0:8000. The default isolation is none (Pi as a child process). OPENAI_BASE_URL on the gateway is the model host, not this API. The client bearer is the model key unless you set OPENAI_API_KEY_OVERWRITE. For Firecracker:

APIPI_RUN_MODE=microvm apipi serve

In a second shell, point a client at http://localhost:8000/v1. agent.model must exist on the model host. The openai package is not an ApiPi dependency:

export OPENAI_API_KEY=dev-token
export OPENAI_BASE_URL=http://localhost:8000/v1
uv run --with openai python examples/openai_sdk.py

A full script is examples/openai_sdk.py. The stream stays open across idle, so that script stops after the first turn outcome.

License

MIT. See LICENSE. Contributions: CONTRIBUTING.md.