Built because the fixes were never run.
We started Korvu after watching AI reviewers propose patch after patch that nobody — human or machine — had ever executed. The tool we wanted reviews the diff, writes the fix, runs your build and tests against it, and only then proposes it. So we built that, in the open.
The bottleneck isn’t writing code. It’s trusting it.
Every engineering team we talked to had the same story. The AI reviewer showed up, the comment volume went up, and trust went down. Suggestions shipped that had never been compiled. Walls of nitpicks buried the one finding that mattered. And the whole thing ran on someone else’s cloud, reading diffs the team had never agreed to send anywhere.
Our bet is that AI review earns trust the same way a colleague does: by showing its work. A fix that arrives with the record of the build and tests that ran against it is a different object from a suggestion — it can be checked, audited, and reverted with confidence. And the engine that produces it should run on infrastructure you control, under a license that keeps it open. None of this is glamorous; all of it is the difference between a tool engineers run on their critical path and a tool they bypass.
Four commitments we make to teams we work with.
Opinionated and specific — the kind of thing we’d want in writing from any vendor on our critical path.
Fixes prove themselves, or they don’t ship.
Verification is the gate in the middle of the loop. A patch that fails your repo’s own build and tests is reported as a finding, never proposed as a commit. We would rather ship fewer fixes than unproven ones.
One engine, one license.
Korvu Core is AGPL-3.0 — the whole loop, not a teaser. No open-core split where verification is the paid bit, no relicensing rug-pull after teams commit. Cloud is a managed layer around the same engine, nothing forked, nothing held back.
Self-host at any size.
No seat gate, no enterprise tier between you and your own deployment. If a capability genuinely can’t exist on self-host (multi-tenant billing, managed credentials), we say so explicitly instead of pretending the gap isn’t there.
Honest surfaces over impressive ones.
Sample data is labeled sample data. In-development capabilities are marked in development. Empty states say so. We don’t publish benchmarks we can’t reproduce or logos we don’t have.
Small now. Built to grow honestly.
One founder, plus AI agents on the unglamorous work. The product decisions, customer calls, and architecture live with a human you can talk to — and we’d rather say that than fake a team page.
Ten years building developer tooling — most recently on a platform team that ran code review through three homegrown bots. Korvu is that platform without the duct tape, and without the part where nobody ran the fix.
Humans decide. Agents draft.
AI agents handle scaffolding, internal tooling, and doc drafts. Architecture, customer conversations, and every line that ships to the engine get human judgment.
- Customer conversations — human
- Architecture decisions — human
- Scaffolding, drafts, tooling — AI-assisted
One founding platform engineer.
Post-GA we’re hiring one founding engineer for the sandbox and verification layer. If making “run the repo’s own tests, fast, in isolation” your whole job sounds fun, we should talk.
Remote-first. Async-friendly. In the open.
Engineering happens across Europe and the US East Coast. The roadmap, the architecture decisions, and the issue tracker live on GitHub — the same place the engine does.
- EUGermany · Portugal · Netherlands. Most of the engineering work happens here.
- US-ENew York · Boston. Partnerships and overlap with West Coast teams.
- GITHUBPublic by default.Issues, discussions, and architecture decision records are open — file a bug where we’ll actually see it.
Two ways in. Both honest.
Self-host the open-source engine today, or join the Cloud waitlist for the managed layer. Same engine either way — that’s the point.
or email [email protected] — a human reads it