GitHub AI Daily · 2026-10-20 (Tuesday)

Picks (5)

1. MedMCP — medical imaging agents choose the "inside the data centre" route

Repo: https://github.com/medmcp/medmcp

An open ecosystem for medical imaging workflows: clinicians describe what they need in natural language, a locally deployed model does the planning, and the actual computation goes to tested tools. The whole stack comes up with a single docker compose command; nothing leaves the hospital network.

2. M4 v0.5.9 — the clinical semantic layer starts versioning skills like software

Repo: https://github.com/hannesill/m4

Released v0.5.9 on 10-19, with changes concentrated in the vitrine-api skill's image examples and a version pin on Vitrine 0.1.5; v0.5.8 had just added the dataset catalog. M4 exposes MIMIC-style clinical datasets through a clinical semantic layer, and now ships that layer per skill version.

3. ai-memory — coding-agent memory is becoming cross-vendor infrastructure

Repo: https://github.com/akitaonrails/ai-memory ｜ +666 stars today, still merging PRs at 07:18

A long-term memory layer written in Rust. Today's merges route OpenCode's Haiku 5.5 to Messages and add MCP workspace query conflict tests. Memory is no longer a feature of one harness but a shared layer across them.

4. HealthClaw Guardrails — someone is filling the gap between FHIR and MCP

Repo: https://github.com/aks129/HealthClawGuardrails

A burst of commits from 05:00 to 07:14 fixed six trust breakpoints found in the 10-18 patient walkthrough, and added documentation for daily message-limit resets, day-2 and day-7 follow-up emails, and a file picker for uploads. This is the PHI, audit and authorisation layer the FHIR-to-MCP path was missing.

5. FoodGuard v1.0.8 — nutrition apps start showing their dataset cards

Repo: https://github.com/qjngbac/FoodGuard

Released v1.0.8 with the matching APK. Fully offline: scan a barcode or photograph the ingredient list, parse on-device, detect additives, check against GB 2760 usage limits and national nutrition standards, then grade A/B/B-/C/D with an itemised deduction ledger (AP/NP/PP).

Trends

1. Medical agents compete on the middle layer, not on capability. MedMCP keeps tools and data inside the hospital, HealthClaw handles PHI, audit and authentication, M4 owns clinical semantics, Hospilot owns observability — four players all doing "everything except the model", all assuming self-hosting.
2. Clinical skills now carry version numbers. M4's 0.5.9 exists solely to ship image assets inside a skill, which means skills are no longer README snippets but software assets that need regression tests. Same line as "skill scale" and "skill personality", now entering maintenance.
3. Context and memory is the most crowded lane in coding agents. ai-memory (cross-harness memory), headroom (compression) and billion-context (long sessions) charted together — as model capability converges, the engineering value concentrates in "what you feed the model".
4. Hardware vendors are absorbing the health data entry point. COROS shipped an official MCP, Zepp a hosted bridge, following community bridges for Garmin and Polar: "one brand, one MCP" is becoming de facto standard and squeezing third-party aggregators.
5. A third route for nutrition: offline, national standards, honest data. Not photo-based estimation, not cloud-based interpretation, but an on-device rules engine plus an explicit statement of data boundaries — for the China market this may be the easiest form to clear review and to earn trust.
6. Official skill and plugin repositories are becoming product lines. Cursor's plugins (13,057 stars) was still merging PRs; Anthropic's claude-plugins-official also charted. Plugin catalogs now need daily maintenance like an IDE or CLI.