Most AI agents I tried wanted an account, a subscription, and my data on someone else's server. I wanted the opposite, so I built Diviso.
It's a Python agent that runs entirely on your own machine. Point it at a local model through Ollama, or plug in OpenAI, Claude, Grok, Gemini, Mistral, DeepSeek or OpenRouter with your own key. No Diviso account, no server of mine in the middle.
What's inside
- ReAct loop + a planner, with hard step and cost caps on every run
- Persistent memory (SQLite + a plain
MEMORY.md) and TF-IDF search over your local files - A small multi-agent team (Researcher / Coder / Reviewer)
- Tools for Python, JavaScript and Bash, file edits, tests, read-only git, web fetch — all behind an approval gate and a workspace sandbox
- MCP client, a guarded scheduler, checkpoint/resume after every step
- CLI + a small web dashboard
The part I'm happiest with
The core is standard-library-only Python, so the same code runs on a GPU PC, an old laptop, and an Android phone under Termux. It detects your hardware and picks a lite / balanced / pro profile with model sizes that actually fit.
Honest limits
It's v0.1.0. No token streaming yet, and the cloud providers are implemented against their documented APIs — the offline test suite (74 tests, CI green) runs on a deterministic mock provider. Local Ollama is the main path I test by hand.
Repo: https://github.com/Neuzant/diviso-agent
Feedback very welcome — especially from anyone who runs it on weird hardware.


