waken-ollama

Ollama Target adapter for Waken — "nginx for AI agents." Routes a waken Event to a local or self-hosted Ollama daemon via the official ollama Python SDK and returns a Response.

Why Ollama fits here particularly well

Every other Target adapter in this ecosystem wraps a cloud API and needs an API key (ANTHROPIC_API_KEY, GEMINI_API_KEY, ...). Ollama needs none of that — it's a daemon that runs on hardware you already control, the same "no external services" idea behind Waken itself: waken keeps all its own state (sessions, scheduled jobs, the retry queue) in one local SQLite file instead of reaching for Docker, Redis, or Postgres. waken-ollama extends that same idea to the model itself — no cloud account, no billing, no secret to manage, just a local process talking to another local process.

Install

pip install waken-ollama

Prerequisite: Ollama must already be installed and running — either locally (ollama serve, or the desktop app, which starts it automatically) or as a self-hosted daemon reachable over the network — with the model you intend to use already pulled:

ollama pull llama3.2

This adapter only talks to an already-running daemon; it doesn't install Ollama, start it, or pull models on your behalf. If the daemon isn't reachable, calls raise a ConnectionError — that's a setup problem to fix on the machine running Waken, not something this adapter papers over.

Usage

from waken import Runtime
from waken_ollama import OllamaAdapter

runtime = Runtime()
runtime.target("ollama", OllamaAdapter(model="llama3.2"))
runtime.run()

waken send ollama "Build tic tac toe." --wait

By default OllamaAdapter talks to http://127.0.0.1:11434, or whatever OLLAMA_HOST is set to in the environment — the ollama SDK's own convention, read automatically. Point it at a remote/self-hosted daemon with host=:

runtime.target("ollama", OllamaAdapter(model="llama3.2", host="http://gpu-box:11434"))

Ollama's /api/chat endpoint has no server-side conversation state, unlike Claude's resume or OpenAI's previous_response_id — the full message history has to be resent on every call. OllamaAdapter keeps a waken session_id -> list of chat messages map internally and grows it in place across turns of the same event.session_id, so a multi-turn conversation picks up where it left off without the caller managing history itself. Extra keyword arguments passed to OllamaAdapter(...) are forwarded to ollama.AsyncClient(...).

Development

git clone https://github.com/WakenHQ/waken-ollama
cd waken-ollama
pip install -e ".[dev]"
pytest

Tests mock ollama.AsyncClient entirely — no local Ollama daemon or network access is needed to run the suite.

License

MIT