LiveDocs

latest-version freshness · contested set (6 fast-moving libs) · captured 2026-07-02
  LiveDocs  ██████████████████████████  6/6
  context7  ░░░░░░░░░░░░░░░░░░░░░░░░░░  0/6
  freshness only · context7 wins coverage · LiveDocs current by construction · measured, not asserted

Measured by evals/look-up/compare_context7.py from a dated snapshot — freshness only, on deliberately fast-moving libs; context7 wins on coverage. Regenerate: python3 evals/look-up/compare_context7.py --emit-readme --write.

A Claude Code plugin for primary-source-first, always-latest documentation.

context7 is a pre-built, periodically-recrawled, lossy vector index. LiveDocs instead auto-discovers each library's canonical machine-readable source on demand and reads it live. It can reach raw primary content and the exact latest version, which a pre-built index can't.

Install

Two slash commands in Claude Code:

/plugin marketplace add PsychQuant/livedocs
/plugin install livedocs@livedocs-marketplace

Or just paste this to your AI and it'll set it up for you:

Install the "livedocs" Claude Code plugin from https://github.com/PsychQuant/livedocs.
Run  /plugin marketplace add PsychQuant/livedocs  then  /plugin install livedocs@livedocs-marketplace

After installing, just ask your usual documentation/version questions — the bundled look-up skill routes each one to the right tool automatically. (LiveDocs is a network-only plugin; its MCP server is signed + notarized and distributed through the marketplace.)

Why this exists

The ecosystem is moving toward machine-discoverable primary sources (llms.txt, package registries, OpenAPI). Live probing of 25 popular docs hosts (2026-06-30) found about 88% already ship llms.txt, and the rest resolve deterministically via package registries. That makes a pre-built index increasingly redundant for the head of the distribution, and it's a source LiveDocs can read verbatim.

Discovery chain (primary-source-first)

1. llms.txt / llms-full.txt. The LLM-designed index/full dump, probed across root, /docs/, and full-variant paths, with a soft-404 guard (a real hit is 200 + text/plain + non-trivial size; many hosts answer 200/404 with an HTML shell).
2. Package registry. npm (registry.npmjs.org/<pkg>/latest), PyPI (pypi.org/pypi/<pkg>/json), and 7 more: the exact latest version plus changelog / repo / docs URLs, deterministically, no scraping. Version pinning supported.
3. Repo. GitHub README / CHANGELOG / releases (raw).
4. Introspection (introspect). Read directly from the machine or installed artifact: an OpenAPI / GraphQL schema, an installed CLI's --help, or an installed R package's version (kind:"r-pkg", read-only). This is the local half of a version check. For a target that has both a web-latest and an installed version, the look-up skill reconciles them and answers from the installed version; the web-latest only gates the upgrade.
5. Fallback. context7 / web, always labeled low-fidelity.

Results are ranked fidelity-first, then freshness. An ETag cache revalidates unchanged sources cheaply without ever serving stale content; it's bounded (LRU + byte budget) so a long session can't grow it without limit.

Every outbound fetch is validated (v0.7.0): an http(s) scheme allowlist plus a host classifier that rejects loopback / link-local / private / metadata targets — re-checked on each redirect hop and against DNS resolution — so a prompt-injection in fetched docs can't steer a request at an internal address. Response bodies are streamed under a size ceiling (bounds the decompressed size, defeating gzip bombs), and returned text is stripped of control / ANSI / bidi characters.

Tools

Tool	Purpose
resolve_source	Ranked primary sources for a library (library+ecosystem and/or docs_url).
fetch_docs	Raw verbatim text of a source URL.
latest_version	Latest released version + changelog/repo, from the registry (9 ecosystems; version pinning on npm/pypi/crates/go/rubygems with a pin_honored signal).
introspect	OpenAPI / GraphQL schema, an installed CLI's --help, an installed R package's version (kind:"r-pkg"), or the effective language-runtime version of the current project (kind:"runtime", Python/Node/Go/Rust/Java/.NET/Swift/R; javascript aliases to node). Read-only.

Architecture

• LiveDocsCore. Dependency-free discovery logic (candidate generation, soft-404 classification, registry parsing, ranking). Network is injected via the HTTPClient protocol, so the engine is unit-tested without a server or the network.
• CheLiveDocsMCP. A thin MCP stdio shell over the engine.

The fuzzy "which library is this?" decision belongs to the calling agent (the look-up skill); these tools take concrete inputs and do deterministic work.

Develop

swift test     # unit tests (pure logic + engine via injected HTTP fakes)
swift build    # builds the MCP executable

CI runs swift build + swift test on every push/PR. scripts/release.sh gates the release on a green swift test, a version-source match against the tag, and Developer ID signing + notarization.

Skill eval (evals/look-up/)

A Python harness that measures whether the look-up skill actually fires a LiveDocs query for varied end-user prompts and answers currently — the trigger reliability the whole product depends on. The oracle is rot-proof by design (version cases fetch the registry at eval time rather than hardcoding a fact), and it's statistical (N runs, rate thresholds).

pip install -r evals/look-up/requirements.txt
python3 -m pytest evals/look-up/tests/     # 60 harness unit tests (no API calls)
python3 evals/look-up/run_eval.py --dry-run
python3 evals/look-up/run_eval.py --runs 5 # live baseline (real `claude -p` calls)

The same directory carries the vs-context7 freshness comparison — a dated, honest head-to-head (registry = neutral ground truth) whose numbers the wiki homepage cites:

python3 evals/look-up/compare_context7.py                # table + tally
python3 evals/look-up/compare_context7.py --verify-live  # warn if the snapshot drifted

See evals/look-up/README.md for the design of both, and the Testing wiki page for the full suite breakdown (204 tests: 144 Swift + 60 Python).

Status: shipped (v0.8.0; engine features below shipped v0.7.0, the skill's look-up rename + explicit invocation landed in v0.8.0). 9-ecosystem registry resolution, version pinning (npm/pypi/crates/go/rubygems), OpenAPI/GraphQL/CLI + installed-R + language-runtime introspection (Python/Node/Go/Rust/Java/.NET/Swift/R, javascript aliased to node, active-toolchain authoritative), bounded ETag revalidation cache, SSRF-guarded + size-capped fetch, the look-up skill (per-question has-local/web-only classification + detect/offer version reconciliation), and a measured vs-context7 freshness comparison cited on the homepage. Signed+notarized release, marketplace distribution.

Design boundary (what LiveDocs is and isn't for): docs/wiki/Primary-Source-Spectrum.md.

Wiki: https://github.com/PsychQuant/livedocs/wiki, mirrored 1:1 from docs/wiki/ via scripts/sync-wiki.sh (edit docs/wiki/, then run the script). Positioning (vs context7): docs/positioning.md.