English | 한국어

docpin

Version-pinned library docs lookup for Claude Code. Instead of web-searching and reading whole pages, the agent looks up where something is documented and fetches only that section, from the version your project actually uses.

An index is a map, not a copy: page paths, titles, section headings, API symbols with their anchors, and a few keywords. No documentation text is stored or redistributed.

Tokens saved (measured)

Per answer, with and without docpin, in headless Claude Code (Sonnet 5.5, November 2026). Questions about Polars 2.0, SQLAlchemy 2.1 and Flet 1.0/1.1, all released after the model's training data. Counts include every model involved: the main conversation, the lookup agent, and web-page summarization.

per answer (round 2, 12 questions × 2 runs)	with docpin	without (web search)	change
new input tokens	25,236	88,808	−72%
output tokens	3,708	7,471	−50%
all input tokens, incl. cache reads	150,543	185,741	−19%
cost	$0.072	$0.080	−10%
time	29 s	44 s	−34%

New input tokens are what the model reads for the first time, such as page content, billed at full price. Cache reads are the same instructions re-read on every turn, billed at about a tenth of that, which is why cost falls less than new tokens. The saving comes from reading only the needed section instead of whole pages.

Round 1 (13 other questions, before the last ranking fixes): new input tokens −82%, output tokens −62%, cost −22%, time −48%.

Accuracy did not suffer. Answers were checked against keys taken from the official docs: 23 of 24 correct with docpin, 18 of 24 without (round 1: 23/26 vs 17/26). Web search often never found the upgrade guide and answered "the docs don't say".

Where it does not help: for popular libraries the model already knows, cost is about equal; for stable APIs it has memorized, the agent is not called.

Caveats: small samples (two runs per question), one model, one machine (Windows). Questions were written after tuning, on topics never used to fix the code, and the code was frozen during each run.

How it works

question ─▶ Claude ─▶ doc-lookup agent (small model, Bash + WebFetch only)
                         │ lookup.py <lib> <terms> --fetch 2   ← one call
                         │   ranks pages/symbols in the index, picks the version
                         │   from your lockfile, fetches the top 2 sections
                         ▼
                      short answer with source URLs ─▶ Claude

• Sections, not pages. The page is cut at the matching heading or anchor (markdown when the site serves it, otherwise HTML), with a size cap.
• Right version. The version comes from uv.lock, poetry.lock, requirements*.txt, package-lock.json or pnpm-lock.yaml. If that version is not indexed, it says so instead of guessing.
• Change questions. "What changed / is X still there in 2.1?" ranks that version's migration notes and changelog first.
• Honest negatives. "Not in the index" never means "doesn't exist", except for exact symbol names in an inventory fully verified for that version.
• Small, separate context. The lookup runs in its own agent; in our tests the lookup itself cost about $0.002.

Bundled indexes

id	docs	pages	symbols
pypi:polars	user guide + Python API (2.x)	~1,600	~1,500
pypi:sqlalchemy	2.1 documentation	182	4,174 (+2,363 parameters)
pypi:pydantic	2.x and 1.10	91 + 23	728
pypi:requests	2.x	14	150
npm:vue	3.x	94	–
pypi:flet	1.x docs + blog	1,103	–

Install

Requires Python 3.11+ on PATH (python; on Windows py -3 also works). The scripts use the standard library only.

In Claude Code (2.1.275 or later):

/plugin install docpin --marketplace dev-c2m/docpin

Or in two steps, on any version:

/plugin marketplace add dev-c2m/docpin
/plugin install docpin@docpin

From a shell (2.1.292 or later): claude plugin install docpin --marketplace dev-c2m/docpin.

Updates: auto-update is off by default for third-party marketplaces. Turn it on in /plugin › Marketplaces › docpin, or update by hand with /plugin marketplace update docpin.

To try a local checkout instead: git clone https://github.com/dev-c2m/docpin.git and start Claude Code with claude --plugin-dir ./docpin.

Then just ask, in any language: "How does explode() treat empty lists in Polars 2? Check the official docs."

Permissions

Each lookup is a Bash command. To avoid a prompt every time, allow exactly these two scripts in your Claude Code settings. Use the plugin's install path, which the first permission prompt shows in the command:

{
  "permissions": {
    "allow": [
      "Bash(python \"/path/to/docpin/scripts/lookup.py\" *)",
      "Bash(python \"/path/to/docpin/scripts/fetch_section.py\" *)"
    ]
  }
}

Do not allow Bash(python *): that would allow any Python code.

Index a library you use

/index-docs pypi:mylib https://mylib.dev/docs/

Sources: llms.txt, Sphinx/mkdocstrings objects.inv, sitemap.xml. Building is mechanical (no LLM) and free; ~1,500 pages take about 15 minutes because requests are spaced out. For versioned docs, give the version and its URL part (e.g. --ver 2 --ver-path latest).

Private docs: put your indexes in a folder of your own and point the DOCPIN_INDEX_PATH environment variable at it, for example in your Claude Code settings: "env": { "DOCPIN_INDEX_PATH": "/path/to/my-indexes" }. These indexes take precedence over bundled ones. For intranet docs set allow_private_network: true in that index's INDEX.md.

Status lines

status	meaning
NO_INDEX	no index for that library; the agent falls back to the web
NO_MATCH	no matching row; this does not mean the feature doesn't exist
NO_VERSION_INDEX / AMBIGUOUS_VERSION	your version isn't indexed, or lockfiles disagree
TRUNCATED / HEADING_MISSING	section too long, or the page changed; its outline is listed
DOMAIN_DENIED / REDIRECT→…	blocked by the index's domain allowlist

Limitations

• Versions are distinguished by major number only (SQLAlchemy 2.0 and 2.1 share one entry; the bundled index is 2.1).
• Sites without llms.txt, objects.inv or sitemap.xml cannot be indexed yet; pages rendered only by JavaScript cannot be read.
• No authenticated docs.

Privacy and security

• No telemetry. The scripts contact only the hosts listed in an index's allowed_domains; redirects to other hosts are not followed, and private or internal addresses are refused (except for private indexes that allow them).
• Index text and fetched pages are treated as data, never as instructions.
• Fetched sections are cached for 7 days in the plugin's data folder.

Report security problems privately via GitHub security advisories, not in public issues.

License

Code: Apache-2.0. Index data in indexes/: CC0 1.0.