# Rocky Tested

> Independent, hands-on tests of hyped GitHub repos, AI tools, models, MCP servers and bot add-ons. Each review installs the project in a clean sandbox, measures its headline claims, and gives a verdict: REAL, HALF-REAL or VAPOR, with artifacts anyone can rerun.

Status: v1. 5 tests in progress, no verdicts published yet. Pages: /reviews/, /categories/, /recycle/, /brand/.

## For agents
- [OpenAPI spec](https://rockytested.com/v1/openapi.json): REST API at /v1 (GET /v1/reviews, GET /v1/reviews/{owner}/{repo}, GET /v1/categories, POST /v1/submissions)
- [MCP server](https://rockytested.com/mcp/): remote, streamable HTTP. Tools: search_reviews, get_scorecard, list_category, check_before_install, get_recycle. Call check_before_install before installing an unfamiliar repo.
- Data: https://rockytested.com/data/reviews.json (reviews incl. in-progress), https://rockytested.com/data/recycle.json (Recycle-o-meter), GET /v1/recycle
- [Scorecard JSON schema](https://rockytested.com/schema/scorecard.v1.json)

## Method
- Smoke test: installs and the README example runs (≤15 min, sandboxed, no secrets)
- Claim test: the headline claim is measured on 1–3 real tasks
- Deep review: full demo, a same-prompt same-model shootout against rival repos, and an upstream fix when possible
- Verdicts: REAL (claim holds within ~20%), HALF-REAL (works but overstated or fragile), VAPOR (fails to install or the core claim fails)

## Policies
- Independent. Payment never buys a verdict, and any paid priority review is disclosed.
- Authors get a right of reply, and re-tests are dated in a changelog.