Most project tools now ship an AI sidebar. You highlight a task, ask a question, get a paragraph back. Useful — but the assistant is usually a layer on top of the product, not a peer that can safely act on live data with your permissions.
Gealo ships two related but different surfaces:
- In-app agent — chat inside a project with persistent conversation sessions (optional
project_id), same permission model as the UI. - MCP server — remote
POST /mcpon your tenant host for Claude Desktop, Cursor, VS Code, and other MCP clients. OAuth consent, auditable tool calls, 66+ tools.
What a registry buys you
An MCP registry is not “more prompts.” It is a typed, permission-filtered API the model can call:
- Task dependencies —
list_task_dependencies,add_task_dependency,remove_task_dependencyagainst real graph edges, not pasted task IDs. - Agent-readable notifications —
list_notifications,get_unread_notification_count, mark-read tools scoped totenant_id(not cross-tenant). - Catalog + native split — 34 native tools in
tools/list; the long tail viasearch_tools→execute_toolso clients are not flooded. - Writes gated —
mcp:readby default; destructive ops preview first (apply_pending_action).
When a chat overlay is enough
If you only need a quick summary of visible text on screen, a sidebar is fine. You do not need OAuth, consent rows, or a tool directory.
When MCP wins
Use MCP when the agent must list sprints, trace blockers, update tasks, or read notifications across a workspace — with the same enforcement as the REST API. That is the “connect your AI to Gealo” moment: not a demo chat, but tools that fail closed when permissions do not match.
See the full tool directory, the comparison page, and the governance spec for scopes and error codes.