AI extensions digest, Oct 23: a codebase map for Claude Code, a read-only Garmin bridge, and an agent reader for long PDFs
Three new open extensions: Vesalius builds a local codebase map Claude Code queries through five MCP tools, garmin-personal-mcp wraps Garmin's newly opened Personal API with a revocable key, and vestigium gives agents addressable sections of long PDFs without a vector database.
AI extensions digest, October 23: a codebase map for Claude Code, a read-only Garmin bridge, and an agent reader for long PDFs
Three new open extensions worth a look today: Vesalius builds a local codebase map that Claude Code can query through five MCP tools, garmin-personal-mcp pipes Garmin's newly opened Personal API into your agent with a revocable key, and vestigium gives agents addressable sections of long PDFs without a vector database. All three appeared on GitHub on October 22, 2026.
Vesalius: a codebase map for Claude Code
Vesalius (haplollc/Vesalius) is a macOS 15+ app, written in Swift, that indexes a codebase entirely on your machine and exposes the result as five MCP tools that Claude Code (or a VS Code extension) can call instead of re-reading files. The pitch is context savings: instead of an agent burning tokens walking a tree, it asks structured questions about the map. The vendor's README reports its own benchmark - 20 questions, where indexed lookups used roughly 113k tokens against 204k for direct reading, with answer quality judged about the same by a blind AI judge. Treat those as the developer's own measurements, not an independent evaluation. One caution for compliance-minded teams: the repository currently has no LICENSE file, so default copyright applies until the authors add one - say "source available", not "open source". 21 stars as of October 23.
garmin-personal-mcp: your watch data, read-only, in any MCP client
Garmin opened an official Personal API in October 2026 with individually issued, revocable keys, and garmin-personal-mcp (Mason-Levyy/garmin-personal-mcp) wraps it as an MCP server - also published on PyPI and listed in the official MCP Registry. The design choices are the right ones for health data: read-only access, no Garmin password anywhere in the flow, and the key can be revoked from Garmin's side at any time. Agents can query sleep, HRV, Body Battery, stress and training load, which makes "morning briefing that knows how I slept" a ten-line configuration instead of a scraping project. MIT licensed (LICENSE file read), Python. 24 stars as of October 23.
vestigium: long-PDF reading with addresses instead of embeddings
vestigium (Yonagi043/vestigium) takes a different bet on agent document reading: skip the vector database entirely. It parses a PDF locally with Docling (OCR included) and gives the agent an addressable structure - sections, pages and tables it can request by reference, like s12, p45 or tab4 - so the model reads deliberately instead of receiving a chunked similarity pile. The authors publish their method alongside the numbers in METHODS.md: their own benchmarks, scored by a blind AI judge, report Haiku gaining 1.2 points over whole-text reading at 0.51 of the cost, 21.2 points over raw PDF input, and Opus answering all 27 questions correctly. Notably, the README states that a human spot check "has not been done yet" - that honesty is a good sign, and it also means the usual caveats about first-party, AI-judged results apply. MIT licensed (LICENSE file read), Python. 11 stars as of October 23.
The short version
Vesalius trades tokens for structure in Claude Code sessions but ships without a license; garmin-personal-mcp is the cleanest possible bridge to Garmin health data, built on Garmin's own new API; vestigium bets that addressable document structure beats embeddings for long PDFs, with self-reported numbers to match. All three are a day old - small stars, early code, worth watching more than adopting blindly.