scoutdocs-mcp

MCP server that fetches and searches current documentation for any package — the latest stable release by default, or one exact version on request. Keeps AI coding agents in sync with current APIs instead of relying on stale training data.

Two ways to run it:

• Local stdio (Python) — installs into Claude Code / Claude Desktop / Cursor, can read your project's manifests.

• Hosted Worker (Cloudflare) — public HTTPS endpoint at /mcp, no install required.

Why

LLMs are trained on a snapshot — the docs they "know" may be months or years old. scoutdocs-mcp gives any MCP client live access to version info, READMEs, and search results across docs sites for packages on PyPI, npm, and crates.io — including exact-version lookups that never fall back to a different release.

Related MCP server: open-context

How it works

Tools

Exact versions

Both doc tools take an optional exact version (local server). The rule is simple: served exactly, or not at all — the latest release is never silently substituted.

flowchart LR
  A["ask: package @ 2.31.0"] --> B{"version exists<br/>on the registry?"}
  B -->|"yes"| C["docs for exactly 2.31.0<br/>+ source URL in the reply"]
  B -->|"no"| D["explicit not-found<br/>no substitution"]

> get_package_docs package="click" version="8.1.7"

The reply says what it is and where it came from:

# click v8.1.7 (python)
Version: exact (8.1.7)
License: BSD-3-Clause
Source: pypi_description (https://pypi.org/pypi/click/8.1.7/json)
Docs: https://click.palletsprojects.com/

---

$ click_
==========
Click is a Python package for creating beautiful command line interfaces…

What that buys you when a version matters:

• Unknown version → explicit "not found". Latest is never substituted.

• version="latest" is refused — pass a real version, or omit version to get latest stable.

• Cache entries are per-version: one version's docs are never served for another.

• On a pinned project, feed the declared_version from detect_project_dependencies into version= to read exactly what the project builds with.

Exact versions run in the local server; the hosted endpoint serves latest stable.

Quickstart

Local (Python stdio)

scoutdocs-mcp is currently a beta (0.2.0b4), so pip and uv need to be told pre-releases are OK:

pip install --pre scoutdocs-mcp                # or
uv tool install --prerelease=allow scoutdocs-mcp

Add to Claude Code's MCP config (~/.claude/claude_code_config.json):

{
  "mcpServers": {
    "scoutdocs": {
      "command": "uvx",
      "args": ["--prerelease", "allow", "--from", "scoutdocs-mcp", "scoutdocs-mcp"]
    }
  }
}

This runs the newest beta without pinning a version; change the --from value to scoutdocs-mcp==0.2.0b4 if you'd rather freeze it. Once scoutdocs-mcp reaches 0.2.0 stable, the --pre / --prerelease flags go away. For Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS), use the same JSON shape.

Hosted (Cloudflare Worker)

Point any MCP client at the public Streamable HTTP endpoint:

https://scoutdocs-mcp.solmonger.workers.dev/mcp

The hosted endpoint is unauthenticated and rate-limited (60 MCP req/min, 10 search req/min per client IP). Search results are capped tighter than local (8 discovered pages × 18k chars / 200k total backstop) but well within Cloudflare free-tier headroom.

Example prompts

> What's the latest version of flask?
> Show me the docs for the serde crate
> Show me the docs for requests 2.31.0 — the exact version my project pins
> Search the httpx docs for "transport"
> What dependencies does this project declare?

Configuration

GitHub token (optional, local only)

Unauthenticated GitHub API allows 60 req/hr. A token (any scope) raises that to 5,000/hr — useful when fetching READMEs in bulk:

export GITHUB_TOKEN=ghp_your_token_here

Cache

• Local stdio: SQLite at ~/.cache/scoutdocs-mcp/cache.db, 24h TTL.

• Hosted Worker: Cloudflare KV, same 24h TTL, scoped per binding.

Search caps

The README is always included as a free first page. The max_pages cap governs how many discovered pages we add on top of it.

Supported ecosystems

If no ecosystem is specified, registries are tried in order.

Repository layout

src/scoutdocs_mcp/    Python stdio server (published as scoutdocs-mcp)
worker/               Cloudflare Worker (TypeScript) for the hosted endpoint
tests/                pytest suite (mocked HTTP)
worker/test/          Vitest suite (Cloudflare workers pool)
docs/RELEASE.md       Release & deployment runbook

Status

Beta (0.2.0b4). API stable; some discovery sources may evolve. Filed issues welcome at https://github.com/eshaanmathakari/scoutdocs-mcp/issues.

License

MIT