Shrinking an agent's macOS tool surface from 84 tools to 19 cut schema context by 73%
bulutarkan · reddit · 2026-09-10
A developer building a local macOS control layer for AI agents found that tool schemas themselves are a costly context tax: exposing all 84 capabilities made every conversation pay even when most tools were irrelevant.
He now uses a "small advertised core + runtime discovery" pattern:
- 19 core tools exposed by default
- 84 capabilities still reachable via tooldiscover + toolinvoke
- Advertised tool-schema context dropped from 16.7k to 4.5k tokens (73%)
For browser control:
- Stable Safari/Chrome tab handles survive tab index changes, so multiple agents can work without fighting over the active tab
- Background DOM/visual observation lets agents act without stealing focus
- A higher-level browser transaction opens, waits, extracts only requested fields, verifies, and closes in one call — returning full DOM to the model is as wasteful as an oversized tool catalog
Open source: mac-mcp. Open question: is "small core + runtime discovery" the right long-term pattern for large action surfaces, or should clients prune full schemas themselves?
More from coding & agent
- Browser Use x Link: agents can now pay with one-time virtual cards in 3 clicks — jeff_weinstein · 2026-09-10
- Solo dev ships 3 Keep features in one day using Amp Code, incl. website monitoring — iannuttall · 2026-09-10
- Smart-DSH Adds Mobile UI and Web Push Notifications to DeepSeek Harness — HankYeomans · 2026-09-10
- Grok-Powered Agent Calls 3 Vendors in 30 Minutes and Books a Week of Beach Gear — jeff_weinstein · 2026-09-10
- Full Clay × LangChain conversation on evaluating millions of monthly agent runs — LangChain · 2026-09-10
- Clay runs millions of agent runs monthly with LangSmith online evaluators, not manual review — LangChain · 2026-09-10