Great coding AI shouldn't be a privilege you rent.
Most of what makes a coding agent good isn't the model — it's the tooling around it. So we build that tooling as open-source engines anyone can use, and price the hosted surfaces to fund the work. Never to lock you in.
The model is becoming a commodity. The tooling around it isn't. Structure-aware edits, grounded research, objective quality gates — that's where coding agents win or lose. So we build it as open source and give it away, because a better-tooled field beats a bigger moat.
Three engines today. Many more coming.
Each is useful on its own and pushes the whole field forward — not just our product. They compose into /code, but the engines are the gift, not a wall around it.
Structure-aware code understanding — fewer tokens, fewer breaks.
Grounds decisions in real, cited sources — not stale model memory.
Turns "is this UI any good?" into an objective pass / fail.
The products that fund the mission.
Cheap, metered, BYOK. The paid surfaces pay for the open work — and they're built on the very same engines, so nothing is hidden.
Search
GAHosted, cited retrieval built on research-mcp — verbatim, source-cited evidence in a token-lean payload.
Open Search →Code
previewAn autonomous, GitHub-native harness that composes the engines — structure-aware via blastguard, grounded by research-mcp. Ships draft PRs, never self-merges.
Join the preview →Design
previewProduction markup gated by design-mcp — accessible, on-system, shippable. Verified, not generated.
Join the preview →Measured, not marketed.
Determinism and token-thrift everywhere, so the cost of great coding keeps falling. Every figure here is reproducible against the live API.
fabricated citations by architecture — every quote is extracted verbatim from a fetched source.
tokens per result vs ~480 for a typical scraped page your agent would otherwise pay to read.
single-digit p99 on cached retrieval, measured at the edge in us-east and eu-west.