How a test heals itself
Record once, replay for free, and repair the test when the website changes, without hiding real bugs. The easy case, explained from the start.
See it work before you install it.
Every run in these videos is real: a real browser, real AI calls, and the costs and timings Saffron reported. Start with the tour, then watch a test heal three times, each change harder than the last.
Episode 0 · Getting started
Saffron in three minutes, from zero to green
Install, write a test in plain language, watch the AI agent record it once in a real browser, review the recording in your editor, then replay it with zero AI calls. Ends with the report and the three result states.
Three changes that break test suites, and what Saffron does about each
A renamed button, a page that disappears, and a step with nothing left to do. In every run the final check passes exactly as written, and nothing changes until you accept the proposal.
Healing at runtime
An A/B test removes a whole page and moves its fields into checkout. The agent takes over in the same browser, works out what the check was for, and finishes the run.
Healing by intent
A step that no longer needs doing. The agent finds proof the goal is already met and records a check for it. Then a real error at payment, which it fails red instead of healing.
One feature, one short video
What a release adds and why it matters, shown on a real run.
Time-travel replay: see what the AI did to your test
Every action shows the real page, a live DOM you can inspect, on a timeline that is your Given, When and Then. The cached attempts that failed, what the agent did instead, the fix proved at zero AI calls, and accept or reject on the page, with secrets masked throughout.
Bring your own AI agent: Claude, Codex, Antigravity or Cursor
Record and heal on the AI subscription you already pay for, with one setting. The same recordings, the same fence on every agent, logins left with each vendor's client, and honest labels: Antigravity is experimental and Cursor is not yet verified live.
Healed in CI, reviewed in your IDE
A heal made in CI no longer dies with the job. The pipeline exports a bundle, the CI runs view in VS Code or JetBrains imports it with every check, and you review the diff, screenshot and trace replay, then accept and commit yourself. No second agent session.
Parallel replay and parallel healing
Replay goes wide with --workers, each scenario in its own browser context, at zero tokens. Recording and healing go wide with --agent-workers, each session in a browser of its own. Outline rows stay together, the cost is said plainly, every run shows where the time went, and shards split a suite across CI machines.
Every browser, every screen, one report
Record once, replay in Chromium, Chrome, Edge, Firefox and WebKit. One command runs the whole browser matrix into one report, and a scenario red in Firefox reads as a cross-browser bug, not a flaky test. Device profiles and viewports put a desktop recording on a phone, where a red is the finding.