Built for the firms cloud-hosted AI can't reach.
FlatClaw started as the answer to a recurring conversation: every firm under a data-locality constraint — legal, healthcare, accounting, finance, government — wanted the same product the frontier labs were shipping, and none of them could buy it. The whole project exists to close that gap with open source.
Skyler Truax
Engineer focused on practical agent infrastructure — runtimes, RBAC, deploy automation, and the unsexy plumbing that makes private-cloud AI actually work in production.
Why FlatClaw
The frontier-lab AI coworker product is genuinely useful — task inboxes, scheduled work, persistent memory, tool access. It just comes with a constraint that disqualifies most of the customers who'd benefit from it most: the data has to leave their tenancy on every turn.
FlatClaw is the same shape, built out of components anyone can audit, deployed onto infrastructure the customer pays for and controls directly. The privacy story is a tcpdump away from being verifiable, not a marketing claim.
How I work
- Working components first — every release ships with end-to-end tests for the features in scope. Silent hangs and unverified claims are blockers.
- Boring substrate, sharp surface — Northflank + established open-source for the spine, attention spent on the product the user actually touches.
- Auditable by default — single image, public license, mechanical privacy proof, honest roadmap.
Apache 2.0, OSI-approved.
FlatClaw itself is Apache 2.0. The Console is MIT. The inference image is public on GHCR. Every line is auditable; the patent grant is explicit.
Read the LICENSE files in the repo for the full terms. If you deploy FlatClaw and find a security issue, see SECURITY.md for responsible disclosure. PRs welcome.