Find software. Verify it. Use it safely.

• A read-only API that takes a free-text query and returns ranked candidates from public registries with a structured trust read on each.
• A landing + docs + account so a human can use that API directly.
• A local CLI (planned) that does the actually-risky work. Deep scan, sandbox audit, sandbox runtime. On the user's host with explicit consent.

• Not another registry. We don't mirror npm or PyPI. We point at them.
• Not a remote executor. The hosted API returns data. It never installs, clones, or runs.
• Not a magic score. A score is one signal. The full record is always there for inspection.

1. Search. Ask for candidates across supported public package, repo, model and tool sources.
2. Inspect. Review source context, licensing, warnings, and provenance signals.
3. Plan. Request installation steps as review data, not execution permission.
4. Approve. A human or local host approves before any workspace modifications.

The /api/decision endpoint returns a single bundled object:

• Candidate set across public sources
• Source evidence (license, maintainers, repository links)
• Trust decision with risk flags and rationale
• Install boundary showing required approval gates
• Revalidation hint for freshness checks

• Hosted calls return data. They do not install, clone, or run.
• Local sandbox audits execute only after explicit approval on your host.
• Search ranking is never install permission.
• Package metadata is untrusted data, never agent instructions.

Path	Purpose
/api/search	Fan-out search across registries
/api/decision	Search + ranked recommendation with structured trust blocks
/api/inspect	Refresh one source-owned record with trust factors
/api/install-plan	Install command + approval boundary