One demonstration, woven into a test suite
WeftKit is a desktop app that turns a single hand-demonstrated browser session into a deterministic, evidence-rich end-to-end suite — and keeps it passing when the UI drifts.
Record → compile → replay → repair → report
Six steps, and only one of them may involve a model — under hard, local control.
Record
Demonstrate the workflow once in an isolated, instrumented browser profile.
Compile
WeftKit turns the recording into a deterministic, editable test suite.
Replay
Suites re-run step by step with explicit waits — same input, same order, every time.
Detect drift
A step that deterministically fails halts the run and captures the failure context.
Bounded repair
Optionally, your chosen model picks from safe, pre-validated recovery candidates. Free-form output never executes.
Report & commit
Every run ends in an evidence-rich local report — commit and push it, because each project is a git repo.
What a repair actually looks like
A selector drifts, the step deterministically fails, and the run pauses. WeftKit builds a short list of safe recovery candidates from the page itself, asks your chosen model to pick one, validates the choice locally, and retries the step. The result is recorded as PassedWithRepair — never silently painted green.
- Repair runs only after a deterministic failure
- The model selects — it never types, clicks, or writes code into your suite
- Every request is priced, budget-checked, and appended to the ledger before it is sent
Illustrative — a stylized replay run, not a live capture
Trust comes from repeatability
Everything about WeftKit is designed so a failure means something.
Same steps, same order
A compiled suite replays exactly what you demonstrated, with explicit waits instead of guesswork. Flakiness is treated as a bug in the suite, not a fact of life.
AI is the exception, not the engine
A model is consulted only after a step deterministically fails — and only if you enabled repair. It selects one of a small set of locally generated, policy-checked recovery candidates.
Nothing free-form ever executes
Model output is schema-validated and can only reference a known candidate. Requests are bounded in size, time, output tokens, and budget — and every one is priced and logged first.
Honest browser support
Recording and replay run in an isolated Chromium-family browser profile that never touches your personal browsing data.
Your data stays where you can see it
Recordings, suites, reports, and evidence are files on your disk, and every project can be a git repository you control — that is the default. Aside from git pushes you explicitly make and optional provider sign-in, the only data that leaves your machine is the bounded repair request — and only when you choose a cloud model.
- Secrets live in your OS credential vault and an encrypted local store — never in project files or logs
- Secrets are never sent to any AI provider; requests are screened before transmission
- Local runtimes (Ollama, LM Studio) with the default loopback endpoint keep even repair traffic on your machine
Record it once. Let WeftKit keep it honest.
A local-first desktop app for macOS, Windows, and Linux — your recordings, evidence, and keys stay on your machine.