Add the docs MCP server and a skill that reaches for it

[abernier]

Gives the plugin its first two components: the docs MCP server, and the skill that makes Claude actually reach for it.

Why both

They are different layers, not alternatives. .mcp.json is the only way to connect https://docs.pmnd.rs/api/mcp; the skill is what decides when to use it. Without the skill the tools are present but the model answers from memory — which is the whole problem, since R3F and drei move faster than any training cutoff.

Why no subagent

A third layer (a docs-lookup agent with its own context) was considered and measured first, against the live server:

payload	cost
docs://{lib}/index	213 tok (r3f) → 1,294 tok (drei)
get_page_content page	1,200 – 4,100 tok, avg ~2,700

A typical lookup is one index plus one page: 3–5k tokens. That sits comfortably in the main context, so isolation would buy little and cost a round-trip plus the agent's blindness to the code being edited — it could report what the docs say, but not that the dpr prop in front of it is stale. The layer composes cleanly later if lookups turn out to be broader in practice; the skill just gains a routing line.

Why the skill insists on reading the index first

Not a token optimization — a correctness rule. Page paths differ per library (/api/hooks for r3f, /learn/guides/... for zustand, /abstractions/... for drei) and get_page_content matches exactly, no normalization. The server's own manifest documents zustand paths under /docs/... that 404, so a model following it lands on Page not found and retries variants.

Coverage is narrower than the enum

The tool advertises 11 libraries. Probing each one, only 4 publish the llms-full.txt dump the server parses:

served	not served (404 on llms-full.txt)
react-three-fiber, drei, zustand, docs	a11y, react-postprocessing, uikit, xr, prai, viverse, leva

For those 7, get_page_content always errors and the index resource returns an empty string. So the skill claims only what it can deliver and sends the rest to WebFetch — a description naming all 11 would fire the skill on questions it cannot answer.

Server-side fix in flight: pmndrs/docs#555. Once it lands, the enum and this skill can widen together as each site ships its dump.

No tests

This went through a test suite and back out again, on purpose. A plugin that ships configuration and prose has almost nothing to assert that isn't a file read back to itself, and the checks that did survive were carrying a package.json, a lockfile, vitest, a tsconfig and a CI job — more machinery than the thing they guarded.

Verifying that docs.pmnd.rs serves what it claims went too: that is the docs site's responsibility, and pmndrs/docs#555 tests it there. A consumer asserting things about its provider only duplicates that, on a slower loop, from a repo with no power to fix what it finds.

So the diff is three files: .mcp.json, the skill, and the README section describing them.

---

🤖 Generated with Claude Code