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. Unknown JSON keys return invalid_request. Known OpenAI fields we have not implemented return not_implemented. See OpenAI compatibility.

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

You need Python 3.13 and a model host URL. Live turns also need the Pi CLI (pi --mode rpc) on PATH. The gateway pins Pi 1.0.0; apipi install can install that binary and, if you pick MicroVM, Firecracker plus a guest image.

apipi serve is the API. It stores data in SQLite at .apipi/apipi.db and binds 0.0.0.0:8000. Pi runs in apipi worker, a separate process. For local development, apipi dev starts both. OPENAI_BASE_URL on the gateway is the model host that Pi calls. Clients send Authorization: Bearer; that value is the model key unless you set OPENAI_API_KEY_OVERWRITE. Worker isolation defaults to none (Pi as a child process). For production, set APIPI_RUN_MODE=microvm on the workers. Several API processes share Postgres.

From PyPI

pip install geki-apipi
apipi install
export OPENAI_BASE_URL=http://your-model-host/v1
apipi dev

apipi dev runs apipi migrate, then starts apipi serve and apipi worker as two child processes. uv add geki-apipi works in a project. The import package and CLI are apipi. S3-compatible artifact storage is pip install "geki-apipi[s3]".

From a git checkout

git clone https://github.com/GEKI-AI/apipi.git
cd apipi
uv sync
uv run apipi install
export OPENAI_BASE_URL=http://your-model-host/v1
uv run apipi dev

Prefix every apipi command with uv run in a checkout.

Client

In a second shell, point a client at http://localhost:8000/v1. agent.model must exist on the model host. Install the openai package yourself if you want the official client:

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

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

Production runs the API and the workers as separate processes, and the workers use microvm isolation:

apipi serve
APIPI_RUN_MODE=microvm apipi worker

More setup is on Install.

License

MIT. See LICENSE. Contributions: CONTRIBUTING.md.