Tests your agent can actually run.
TRIF captures an app's testing knowledge once, as portable human-readable recipes — so any AI agent, engine, or human runs a test from a tiny slice of context, grades it against an explicit oracle, and tells a real bug apart from a stale assumption. Open format. Zero-dependency kit. Runs from your CLI or straight from an agent over MCP.
Agents re-learn every app, every run.
A natural-language browser agent can test an app it has never seen — but it pays two costs on every run, and neither gets cheaper with repetition.
01 Re-exploration
The agent re-derives navigation, selectors, and timing from scratch each run. Knowledge from run N evaporates before run N+1 — so cost never amortizes.
02 Verdict conflation
When a step fails, the agent can't tell "the app has a bug" from "the app changed and my assumption is stale." It either cries wolf or silently adapts to a regression.
Explore once. Write a cookbook. Run at O(1-test) cost.
A cookbook (<app>.trif/) is a versioned artifact: recipes, anchors, an app map, and a healing ledger.
Any executor runs one recipe from ~1.5k tokens of context and returns one of four honest verdicts — drift detected and named, never silently absorbed.
Use it however your agent lives.
One artifact, four consumption paths — each a thin wrapper over the same files, results flowing back into the same schema.
A cookbook's router is a valid SKILL.md. Drop it into Claude Code, Codex, or Cursor and the agent auto-loads the workflow.
Run test-harness-kit mcp and any MCP agent drives the whole loop as tools — read a recipe, run it, get a verdict. Setup →
npx test-harness-kit — for CI, cron, and any shell-capable agent on any model. Zero dependencies.
It's just markdown + JSON. Read a recipe and execute it by hand — the format floor that keeps every wrapper thin.
The map, the evidence, and the report — in your hands.
A cookbook shouldn't be a black box. Browse the app's structure, watch what a run did, and read exactly why each recipe passed or failed.
Reviewable IA map
Surfaces × anchors × recipes, color-coded by freshness, with a detail drawer.
Open the live map →
Detailed run report
Per recipe: verdict, step trace, oracle checks, timing, and evidence.
Open a real report →
Native evidence
Every run leaves a Playwright trace, video, console + network logs, and screenshots.
How it works →Pay to learn the app. Don't pay to re-run it.
A recipe that passes on the intent lane graduates to a compiled script binding. The next run executes it deterministically — no model, no API key, $0 — and on drift it falls back to the intent lane and heals. A binding failure is a signal, never a verdict.
Intent lane
A recipe runs via the agent/engine against the live app, graded by its oracle.
Graduate
A PASS on a deterministic-ready recipe compiles to a Playwright (and Katalon) binding.
Run at $0
Fresh bindings run with no model. On miss/stale/failure → back to intent + heal.
A neutral benchmark — reported honestly.
"WebVoyager scores task completion; TestVoyager scores whether a tool can test." Any tool enters via
{url, intent} → verdict — TRIF not required, so agents and vendors enter as peers. Results are pre-registered and published whichever way they point.
The benchmark →