FAQ

FAQ

Does it need an internet connection?

No. With Ollama and a local model, nothing leaves your machine. Cloud providers are optional.

Does it need a GPU?

No. A 7–8B model runs on CPU — slower, not broken. A GPU is used automatically if present.

Is my data sent anywhere?

Not unless you configure a cloud provider. There is no telemetry, no account, and no phone-home.

How is this different from a normal agent framework?

Most frameworks give you agents that talk. This one is built around memory that outlives the session: graded predictions, tracked reputation, preserved dissent, and a constitution that rejects output rather than advising it. The agents sit on top of that.

What happens if I delete the database?

The House forgets everything it has learned — every graded prediction, reputation estimate, and recorded belief change. The system still runs; it starts over. See Institutional Memory.

Why does it reject my forecast?

Because it lacks an invalidation condition. An unfalsifiable prediction can never be graded, so it can never teach anything. See Governance Engine.

Can I use it with GPT / Claude / Gemini?

Yes — ten cloud providers are supported through one adapter, and any OpenAI-compatible server works too. See Choosing a Model.

Can I run it on a server?

Yes, but port 8766 has no authentication and the API can run shell commands. Keep it on loopback and front it with an authenticating proxy. See Security Model.

Is it safe to let it run tools?

It is gated, not sandboxed. The GPS-2 gate is deny-by-default and irreversible actions need a human gate — leave those on, and point it at a workspace you are willing to lose. Read Security Model before enabling execution.

Why fourteen agents? Isn't that overkill?

The count matters less than the distinctness. A council of identical agents is one agent with extra latency. The Skeptic having a veto, and dissent being preserved, are what make it a council. See The Council.

Does it work on macOS?

It should — the Linux code path is POSIX. But macOS is not in CI, so it is listed as untested rather than claimed. Reports welcome.

What licence?

Apache-2.0 — OSI-approved, express patent grant, free commercially at any revenue.

How do I report a bug or a vulnerability?

Bugs: issues. Security: advisories — privately first.

Is this production-ready?

It runs, it is tested (601 tests, CI on Ubuntu and Windows across three Python versions), and it is used daily by its author. But it is a single-maintainer project with no sandboxing and no API authentication. Evaluate it as such, and read Security Model before pointing it at anything you care about.