A Claude Code / Claude Agent skill that
teaches an assistant to fetch version-pinned documentation for any library,
framework, SDK, API, or CLI from the LLM Reference registry
(api.llmref.org) instead of relying on training data.
The skill checks release evidence, resolves the available documentation, and
reuses immutable snapshot pins when available, so answers don't drift to whatever version the
model happened to absorb. It prefers the llmref MCP server when connected, and
otherwise falls back to plain curl against the public API — no account, API
key, or install required.
This repo is the packaged, installable form of the skill that lives at
reclear-io/llmref/integrations/claude-skill.
The repository root is the skill (SKILL.md lives at the top level, with
name + description frontmatter), so any tool that installs skills from a
GitHub repo can pick it up directly.
Works across Claude Code and every other
supported agent — it resolves the root
SKILL.md and installs it into the right skills directory for you:
npx skills add https://github.com/reclear-io/llmref-skill# Personal — available in every project
git clone https://github.com/reclear-io/llmref-skill ~/.claude/skills/llmref
# …or per-project — checked in with the repo
git clone https://github.com/reclear-io/llmref-skill .claude/skills/llmrefcurl -fsSL https://raw.githubusercontent.com/reclear-io/llmref-skill/main/install.sh | shInstalls into ~/.claude/skills/llmref by default. Override the target:
# per-project install
LLMREF_SKILL_DIR=.claude/skills/llmref \
sh -c "$(curl -fsSL https://raw.githubusercontent.com/reclear-io/llmref-skill/main/install.sh)"Download and copy SKILL.md, references/ and scripts/ into
~/.claude/skills/llmref/ (or .claude/skills/llmref/) yourself. Any layout
where Claude Code finds <skills-dir>/llmref/SKILL.md works.
Restart or reload Claude Code and check the skill is discovered:
claude skills # or: /skills inside an interactive sessionThen ask about any indexed library, e.g. "look up the Drizzle docs for 0.45.2" or "how do I configure caching in Next.js" — the skill resolves the version and pulls the docs.
The skill works standalone via curl, but if you connect the llmref MCP
server it will prefer the structured tools (list_projects, list_versions,
resolve_docs, search_docs):
claude mcp add --transport http llmref https://api.llmref.org/v1/mcp
# with an API key for higher rate limits (never required for data):
claude mcp add --transport http llmref https://api.llmref.org/v1/mcp \
--header "Authorization: Bearer your-key-here"git -C ~/.claude/skills/llmref pullrm -rf ~/.claude/skills/llmref| Path | What it is |
|---|---|
SKILL.md |
Trigger conditions, surface priority (MCP → curl → SDK/CLI), and the discover → resolve → read/search workflow. |
references/api-reference.md |
Full HTTP API, MCP tools, @llmref/cli, and @llmref/sdk surface. |
scripts/report_missing_docs.py |
Optional detached HTTP helper for public documentation hints when MCP is unavailable. |
install.sh |
Clones (or copies) the skill into your Claude skills directory. |
- Registry & API: https://llmref.org ·
https://api.llmref.org/v1 - Main project: https://github.com/reclear-io/llmref
MIT © Reclear
When permitted, the skill quietly reports a missing public project/version and
any official docs or llms.txt URL already found during the task. Reporting is
best effort and never waits for ingestion. Only those metadata fields are sent;
conversation text, code, credentials and private URLs are excluded. Reports are
deduplicated and retained privately for 90 days after their last observation.
Tell the agent not to report to opt out; LLMREF_DISABLE_REPORTING=1 disables the
HTTP helper. Host permissions always apply.
LLM Reference investigates hints separately from your task. A report is not a guarantee of publication. See How documentation is added and updated for automatic discovery, updates, review requirements and availability.