The judgment layer for coding agents
Wellworn is built by a founder who spent a year shipping products with coding agents and kept paying for the same mistakes: a library the agent remembered from 2024, a version that no longer existed, a UI that looked like every other generated page.
Documentation tools solved half of the problem. They tell an agent how to call something it already chose. Nothing told it what to choose, or what breaks at the version in the lockfile. Sonatype measured that 27.76% of AI dependency upgrade suggestions point to versions that are non-existent, deprecated, or unsafe (Making AI Work Safely, 2026). That is the gap.
A verdict is one question, one pick, one alternative and the condition under which it wins, what to avoid, the traps at a version, the skill to follow, and the sources with the date they were read. Every verdict expires. A nightly job re-reads the package registries, and when a major version moves, the verdict is served with a RECHECK flag until a person re-verifies it.
The MCP server, the Claude Code plugin, the skills, and the verdict corpus are folders of one public repository, WellWorn-dev/wellworn. The verdict format is CC0 so other corpora can interoperate.
Wellworn runs on its own servers behind Cloudflare. No prompts or code are stored; the usage log keeps the tool name, the scope type, tokens out, and latency.
Write to support@wellworn.dev.