Make Capacities agent-native: give AI agents full access to the workspace

Make Capacities agent-native: give AI agents full access to the workspace

The MCP server is a great start — semantic search, OAuth, creating typed objects. But agents are stuck on single objects, and it's slow. Five additions would fix that.

Why this matters: I want my workspace to be an agent's working memory. Project decisions, sources, and notes live in Capacities; my agent reads them, reasons across them, and writes results back as proper objects. No copy-paste, and context that survives between sessions.

The blocker, straight from the docs:

"retrieving an entire set of objects (e.g., all objects of a specific type) is not officially supported. Prompts requiring analysis or counting across all objects of a type are also not supported."

So "review all my project notes" can't be answered at all.

Asks

1. Enumeration — GET /objects?structureId=…&cursor=…, no required query. Already on the roadmap; the single biggest unlock.
2. Incremental sync — an updatedSince parameter and updatedAt in responses. One param, one field, and agents stop re-crawling everything.
3. Rate limits for bulk reads — 30 req/60s is an interactive budget. Either a batch endpoint (GET /objects?ids=[…]) or a separate read-only quota.
4. MCP ↔ REST parity (most important to me) — MCP has semantic search with snippets, REST returns titles only. REST has delete, collections, and media; MCP has none. Anything a human can do in the UI, an agent should be able to do through MCP — at speed. This decides whether an agent is a collaborator or a toy.
5. ACP support — Agent Client Protocollets Capacities Chat act as a frontend for an external agent (Claude Code, Codex, Gemini CLI) running locally: Capacities ↔ ACP ↔ your agent ↔ MCP ↔ workspace. BYOK brings your own model; ACP brings your own agent — with its own memory, tools, and access to your machine. Not a request to embed a terminal.

Items 1–4 are incremental API work. Item 5 would make Capacities the first serious agent-native knowledge workspace.