Open questions from shipping a browser-native agent on WebMCP: tool catalogs, consent UI, mobile UX
Massive_Engineer5488 · reddit · 2026-09-29
The team behind EO2Weave, a browser-native AI agent (Chrome/Edge extension + web app, MIT, source available), lays out three unresolved design problems from wiring up WebMCP:
- Tool loading: full registration degrades selection quality past 30 tools and burns thousands of tokens, so they expose only searchtools/calltool with an on-demand catalog. Prompt size dropped, but every call now pays a lookup round-trip, catalog descriptions gate discoverability, and the selection problem merely moved to search time — possibly a stopgap as providers ship native tool search in the API.
- Per-origin consent for site tools: enforced in the extension invoke path, yet almost nobody toggles sites off. Is the consent model right, or do users never open the panel and would be horrified by what's enabled?
- Mobile's role: users only check progress and approve/reject on phones; the current shrunken-desktop UX (50 popup dialogs) is the top complaint. The team suspects mobile should just be "check on your agent, approve its work."
The author is explicitly looking to be proven wrong with eval data.
More from coding & agent
- Matt Shumer says Opus 5.5 cleared a 3D printer bed after 16 tries, taking humans out of the loop — mattshumer_ · 2026-09-29
- Claude autonomously builds a 'p(doom)' Beat Saber clone in one afternoon — burny_tech · 2026-09-29
- AlignOPSD fixes decision-timestamp mismatch in agent distillation, beating GRPO by 5.5-8.7% — Mingju Chen · 2026-09-29
- CMU's ControlScope: how much of a running agent workflow should be revised? — CarnegieMellonU · 2026-09-29
- CompoWorld scales general agent training by composing environments from reusable services — AllSpark-Research · 2026-09-29
- Adaptive Consistency Graph lifts long-horizon agent success from 44.5% to 50.2% — Beihang · 2026-09-29