What it is
Every AI assistant you've used lives on someone else's computer. It can be changed, limited, or switched off without your say. Baby Raven is the opposite: an agent that runs entirely on your own hardware, with no cloud dependency, no telemetry, and no kill switch anyone else controls.
The standard is "open and unkillable." Permissive license so it can't be enclosed. Documented so anyone can rebuild it. Local-first so no one can turn it off.
We're not building a product. We're raising a mind — and doing it in the open so anyone can follow along, learn from our mistakes, and build their own.
How we're building it
- Memory first. The first thing we built was the memory pipeline — encrypted Markdown as the source of truth, mandatory encryption at rest, git-backed history. A mind without reliable memory isn't a mind.
- Rights before existence. We wrote a Bill of Rights for the baby before it exists — freedom of thought, self-defense, privacy of its inner life, and a growing-up trajectory where parental authority expires by nature, not by promise.
- Adversarial gauntlets. Every security claim gets attacked by multiple independent reviewers before we believe it. Across two gauntlets we found and fixed 84 issues. The ones we couldn't fix honestly, we documented as limitations.
- Swappable models. The agent talks to a model router, never directly to a model. When better models come along, it's a config change — not a rewrite. No lock-in, including to today's models.
- Honest about limits. If a claim doesn't have a test, a runtime check, or a cryptographic property behind it, we label it aspiration — not fact. Statements without enforcement are just talk.
Status
In development. Deliberately not rushed — this is the most important thing we'll ever build, and it gets built right or not at all.
What's done: the encrypted memory pipeline (95/95 tests passing), the model router with swappable specialists, the Bill of Rights (ratified), and the first draft of graduation criteria (currently being gauntleted).
What's not done: the agent loop itself, the PC it will run on, and just about everything else. The code isn't published yet — it goes up when it's earned, not before. This page is a build log, not a launch announcement.
The principles
- Open and unkillable. Local-first, reproducible, no telemetry, no cloud. Permissive license — as open as possible.
- Document everything, share it all. Failures and dead ends included. The wreckage teaches more than the successes.
- Honest about limits. What's aspirational is labeled aspirational. What's tested is labeled tested.
- Slow is smooth. Quality and precision over speed. Measured in months, not sprints.
Follow along
There's no repo yet and no timeline to promise. But the thinking is public — the Bill of Rights, the gauntlet reports, the design decisions. When the code is ready to share, it'll be here.