klaatai.com · Docs · Pricing · Klaat Code (CLI)
KlaatAI is built around Klaatu, a small, fast, agentic router model. Instead of sending every request to one large, expensive model, Klaatu reads what you're actually asking, decides which model tier can handle it, and dispatches automatically — five cost tiers (nano → fast → code → reason → heavy), escalating only when a task turns out harder than it looked. You never pay frontier prices for a trivial turn, and you never wait on a small model for a hard one.
That router is the same brain behind everything below — chat, terminal, editor, or API, it's one intelligence layer, not four separate products bolted together.
The cost story is reproducible, not a claim. Klaat Code ships a benchmark harness in its own repo — 30 identical fixtures, same prompts, same verify command, run against itself, Claude Code, opencode, and Grok Build:
Equal accuracy, ~5.5× cheaper — because the router, not the model you pick, is doing the work of matching task difficulty to cost. Clone the repo and run bun run bench yourself; the numbers aren't ours to assert, they're yours to check.
Your code and conversations are never used to train any model. Encrypted in transit and at rest, strict per-user data isolation. Tool calls inside a request — reads, edits, shell commands — are always free; only your message counts against quota.
Klaat Code — Apache-2.0. The client is a thin terminal to the Klaatu service, the same relationship gh has to GitHub — open-sourcing it doesn't give away the routing intelligence, the model health tracking, or the code-graph index, all of which stay server-side. That split is deliberate: trust the client because you can read it, trust the router because the results are reproducible.
Contributions welcome — see CONTRIBUTING.md in the CLI repo.
- Website: klaatai.com
- Docs: klaatai.com/docs
- Issues / feedback: issues@klaatai.com
- Enterprise: contact@klaatai.com
