Start here
What Wellworn answers, the one-line install, the eight tools, and the cap on each tier.
Wellworn is an MCP server that answers the question a documentation tool cannot: which stack, library, skill, or design system your agent should use, what breaks at the version in your lockfile, and who checked it. Each answer is a verdict, which is one question with one pick, one alternative and the condition under which it wins, what to avoid, the traps at a version, a skill to follow, and the sources with the date a person read them. Every verdict carries an expiry date, and a nightly job re-reads the package registries, so a verdict whose pinned version has moved is served with a RECHECK flag until a reviewer looks at it again.
The read tools work with no key at all, at 60 calls a day per IP, so an agent can use the server before anyone creates an account. The corpus, the plugin, the skills, and the verdict format live in one public repository, WellWorn-dev/wellworn.
Install
One line in Claude Code:
claude mcp add --transport http wellworn https://mcp.wellworn.dev/mcpVerified 2026-09-08 against code.claude.com
Every other agent is a few lines of JSON or TOML. Each page has the exact file and shape, checked against the client's own docs on the date shown.
The endpoint speaks Streamable HTTP at https://mcp.wellworn.dev/mcp. It is stateless: no session id, one request, one response. Any client that can reach a remote HTTP MCP server can use it, and clients that cannot reach one directly can bridge with npx mcp-remote https://mcp.wellworn.dev/mcp.
The eight tools
| Tool | Arguments | Answers |
|---|---|---|
recommend | task, constraints[] | Which stack, library, or service to use for a task. |
compare | options[] (2 to 4), context | How two to four candidates differ, and when each one wins. |
alternatives | to, kind | Open-source or self-hosted replacements for a SaaS or library. |
traps | library, version, id | Known breakages at a version: symptom, severity, fix. |
skill | slug | The SKILL.md procedure for a verdict, and its install command. |
design | screen, constraints[] | Design system, tokens, layout rules, and anti-patterns for a screen. |
docs | library, topic | Official docs URL, llms.txt, and the latest version. |
submit_trap | library, version, symptom, fix, evidence_url | Reports a breakage you hit. Needs a key. |
The tools reference has the full argument list and an example response for each one.
Limits
| Tier | Cap | Window | Key |
|---|---|---|---|
| No key | 60 calls | per day, per IP | none |
| Free key | 2,000 calls | per month, per organization | free |
| Pro | 20,000 calls | per month, per organization | $12 a month |
| Team | 100,000 calls | per month, pooled across the organization | $49 a month |
Caps are per organization, not per key, so extra keys never multiply an allowance. Paid plans can run past the cap on prepaid credits: 25,000 calls for $10, 150,000 for $50, valid twelve months, consumed only after the plan cap. Every read tool works on the no-key tier; submit_trap is the one tool that requires a key. At the cap a tool returns a single line naming the reset time, so the agent keeps working instead of erroring.