AgentsRoom-MCP currently declares its full tool catalogue (100+ tools across backlog, memory, notes, agents, teams, commands, ssh, db, kube, settings, browser, QA...) to every agent console. Measured on a Claude Code chief-of-staff agent: roughly 22 KB of tool schemas are carried every turn, most of them never called in that session. Claude Code mitigates part of it with deferred loading (ToolSearch), but other CLIs load the whole catalogue up front, and even on Claude Code the catalogue listing itself is repeated.
Proposal: let the MCP server expose a small core set by default (capabilities_get, the families the agent's role actually uses) and fetch the remaining schemas on demand, e.g. a tools_search/tools_load pair or MCP tool-list pagination by family, driven by the agent role or a per-project setting. The desktop already knows each agent's role, so the default set could be derived from it.
If this is already planned, a pointer to the existing ticket is enough.
AgentsRoom diagnostics
- App version: 1.223.1
- OS: darwin 27.0.0 (arm64)
- Agent CLI: claude
- App language: system
Problem to solve
Every agent turn carries ~22 KB of AgentsRoom tool schemas it does not use; on long coordination sessions this is a meaningful share of context and of token spend, and on non-Claude CLIs there is no deferred loading at all.
Proposed behavior
Default to a role-derived core tool set and load the other families on demand (a search/load tool pair or family-scoped tool listing), with a project setting to opt out.
Who benefits
Anyone running many long-lived agents (coordinators, reviewers) across CLIs.
Acceptance criteria
Smaller per-turn context, fewer tokens per session, no loss of capability.