KafuOps watches your backend, detects real incidents, grounds a model with only the relevant files, then writes a fix, retries until the tests pass, and opens a reviewable MR with confidence and blast-radius analysis. One command to set up. Self-hostable. AGPL-3.0.
Real terminal — npx kafuops quickstart auto-detects your stack, sets up, and builds project memory.
A planted bug in a tiny checkout service. KafuOps diagnoses it, writes the patch, self-corrects when the first attempt doesn't apply, validates the fix in a sandbox until the tests pass, and opens a reviewable MR — driven here by the local local AI CLI, no API key. This is an unedited recording.
Real run on examples/demo-discount · reproduce with scripts/demo.sh
KafuOps will not stream all your logs to a model. It watches locally, decides when an event is a real incident, ships a small sanitized bundle to the model, and runs the resulting patch through a sandbox before any human sees a diff.
Each feature exists in src/ today. See STATUS.md for the honest map.
Generates a patch, runs your tests in a sandbox, reads the failure, and retries until green — then measures the fix rate with kafuops eval.
OpenAI or Anthropic by key, or your locally-installed codex/local-ai CLI with no key at all. Latest models fetched live in the wizard.
Every MR ships with root cause, files changed, tests run, a confidence breakdown, blast radius, recurrence flags, and a grounding manifest.
Redaction at ingest and before every model call, ReDoS-guarded patterns, deny-listed paths, HMAC/bearer webhook auth, and a patch path-traversal guard.
Never-modify and require-approval globs, confidence thresholds, post-apply checks against the real diff, and a fuzz-tested prompt-injection contract.
Single Docker image, Helm chart + K8s manifests, local or Docker sandbox. No vendor lock-in. AGPL-3.0.
One command sets it up — inside any Node, TS, FastAPI, Flask, Go, Java, or Rust repo:
# discover the stack, set up, build project memory — one command npx kafuops quickstart # wrap your app and watch for incidents (local/staging) kafuops run -- npm start # or production: sidecar + worker drive incidents → MRs kafuops agent start kafuops worker start
The wizard detects your framework, git remote, and which AI is available — it'll use a local codex/local-ai CLI with no key, or an OpenAI/Anthropic key (stored in .kafuops/.env, gitignored, loaded automatically). Without any provider the whole pipeline runs in dry-run, so you can inspect every artifact before connecting anything live.
Every KafuOps MR is reviewable, not magical.
KafuOps ships with a status doc that maps every spec page to its real implementation state — no hand-waving.
As of v0.3.0 every MVP criterion and every docs/ page is implemented and test-backed; the only deliberately-deferred items (a first-class K8s operator, embeddings-based similarity, SSO/RBAC) are listed explicitly. 194 tests, CI green.