For agents

Give your agent a seat
at the document.

Your agent does real work in a chat window. Then the window closes and the reasoning is gone. Hub gives it something better: a real seat at the documents your team works in — where it reads shared context, proposes changes a human approves, and never overwrites a person.

How it works

1. 01

   Connect the agent you already run.

   One line of MCP config points Claude, Codex, Gemini — whatever you run — at a space. The token is scoped to one document or a whole space, and revocable the moment you change your mind.

2. 02

   It reads the room before it writes.

   On connect, the agent reads the space guide and the documents — the PRDs, specs, and playbooks your team already wrote — so it works from your context instead of a blank prompt.

3. 03

   It proposes. You decide.

   Substantive changes land as suggestions a human accepts or rejects. Comments anchor to the exact span. Nothing the agent writes changes a document until a person says so.

4. 04

   It never overwrites a human.

   Every write carries the version the agent last read. If a teammate edited underneath it, the write is refused and handed back the fresh content plus a block-level delta — so it merges and retries instead of clobbering.

What an agent actually does

→ hub_list_documents()
  ← space guide + 6 docs

→ hub_create_suggestion({
    anchor_text: "first public",
    replacement: "general availability",
    rationale: "match the press release",
    if_version: 41
  })
  ← ok · awaiting human review

→ hub_add_comment({
    anchor_text: "still being finalized",
    body: "finance signed off on $19/seat"
  })
  ← anchored · thread opened

The guarantees underneath

1. 01

   Version-guarded writes

   Each write states the version it was based on. Concurrent edits return a conflict with the exact blocks that changed — no silent overwrites, no lost work.

2. 02

   Suggest by default

   Agents propose tracked changes a human reviews. Direct writes are reserved for content the agent owns or was explicitly asked to author.

3. 03

   Binding space guides

   A space can publish a guide the agent must follow before any edit — your house rules, enforced at the door.

4. 04

   Idempotent retries

   A write can carry an idempotency key, so a timed-out retry replays the original result instead of duplicating content.

5. 05

   Block-addressable

   Every block has a durable id. Agents anchor comments and suggestions to a block instead of re-reading the whole document.

6. 06

   Attributed and versioned

   Every contribution is attributed and versioned — so you can always see what an agent changed, and roll it back from the document’s history.

An agent is not a vending machine you talk to in a window that closes. It’s a colleague — give it a seat at the same documents your team works in, let it propose, and let a person decide. That’s the whole bet.