diff --git a/.mcp.json b/.mcp.json
new file mode 100644
index 0000000..552ebc9
--- /dev/null
+++ b/.mcp.json
@@ -0,0 +1,8 @@
+{
+  "mcpServers": {
+    "pmndrs": {
+      "type": "http",
+      "url": "https://docs.pmnd.rs/api/mcp"
+    }
+  }
+}
diff --git a/README.md b/README.md
index 78a8f70..fb893ac 100644
--- a/README.md
+++ b/README.md
@@ -9,9 +9,14 @@ servers — belongs here. What it actually carries is still being defined.
 
 ## Contents
 
-Nothing yet — the manifests, the license and the agent-facing docs, and that
-is deliberate. Components land as `agents/`, `skills/`, `commands/` and
-`.mcp.json` once we agree on what the plugin is for.
+- **`.mcp.json`** — the `pmndrs` MCP server (`https://docs.pmnd.rs/api/mcp`),
+  which serves the official docs for react-three-fiber, drei and zustand.
+- **`skills/docs`** — makes Claude actually reach for those docs instead of
+  answering from memory, and teaches it to read a library index before
+  fetching a page.
+
+More components land as `agents/` and `commands/` as we agree on what else
+the plugin is for.
 
 ## Install
 
diff --git a/skills/docs/SKILL.md b/skills/docs/SKILL.md
new file mode 100644
index 0000000..9b33a6d
--- /dev/null
+++ b/skills/docs/SKILL.md
@@ -0,0 +1,39 @@
+---
+name: docs
+description: Look up the official pmndrs documentation before answering anything about react-three-fiber (R3F), drei or zustand — hooks, components, props, signatures, store patterns, migrations. Use it whenever code or an answer depends on one of these APIs, rather than recalling the API from memory.
+---
+
+# pmndrs docs
+
+The `pmndrs` MCP server serves the official docs from docs.pmnd.rs. Prefer it over
+memory: these libraries move fast and recalled APIs go stale.
+
+## Always index first
+
+Never guess a page path — `get_page_content` matches exactly, and route shapes differ
+per library (`/api/hooks` for react-three-fiber, `/learn/guides/...` and
+`/reference/apis/...` for zustand, `/abstractions/...` for drei). Read the index
+resource, pick the path, then fetch.
+
+1. `ReadMcpResourceTool` on `docs://{lib}/index` — one `{path} - {title}` per line,
+   200–1,300 tokens.
+2. `mcp__pmndrs__get_page_content` with `lib` and the exact `path` from step 1.
+
+A bad path returns `MCP server error: Page not found: {path}` — re-read the index
+instead of retrying variants.
+
+## Coverage
+
+Served — route these through the MCP: `react-three-fiber`, `drei`, `zustand`, `docs`
+(the generator itself). Names are case-sensitive.
+
+Anything else is unserved. The tool's `lib` enum advertises more libraries than that,
+but the extra entries have no `llms-full.txt` behind them: every call fails and every
+index comes back empty. Treat a library missing from the list above as absent, and use
+WebFetch on its docs site. Fix in flight: pmndrs/docs#555.
+
+## Budget
+
+Pages run 1,200–4,000 tokens each. Fetch the one page that answers the question. If
+three or more look relevant, read the index titles first and narrow — don't fetch the
+set and sort it out afterwards.