Milestones
Concrete things codelazr actually does today, not a roadmap. Each one below is shipped code, not a plan. See /stats for the real numbers behind them.
Code auto-review
A second, cheaper model reads every diff before it commits and flags anything that looks off, never blocking, just warning. After a commit lands, a separate check reads the full repo file tree against the original task, so an incomplete build (missing files, not just bad code) gets caught too, not just what changed this turn.
Bounded self-fix loops
When a build fails CI, when a deploy fails, or when the post-commit review finds a real gap, codelazr hands the actual error (or concern) straight back to the model as a follow-up turn and lets it fix its own mistake, capped at a fixed number of attempts so it never spins forever. A test-then-improve cycle that runs itself, not something you have to copy-paste error logs into.
Cheap models, frontier-shaped results
The whole premise: weaker, cheaper OpenRouter models, nudged by the mechanisms above, can land results that would otherwise need an expensive frontier model. The SaaS landing site case study runs the exact same prompt through many different budget and free models and shows every attempt, including the ones that failed, so this isn't just a claim.
Weak-model failure recovery
Free and small models fail in specific, predictable ways: they say "done" with an unfinished to-do list, they return an empty response, they say they'll ask a question and then don't. codelazr recognizes each of these and offers (or, in Auto mode, automatically fires) the exact right follow-up, so the task actually finishes instead of quietly stalling on you.
GitHub-native, not a sandbox
Every commit, PR, and deploy status shown in codelazr is a real GitHub action against your real repo, read back live through GitHub's own APIs (CI checks, deployments, build logs), not a simulated preview. What you see committed is what actually landed.
More on why codelazr exists and how it's built on /launching-librex.