For your agents

Your notebook speaks MCP.

Point Claude Code, Codex, or any MCP client at a section. It works from your actual notes: the same files, the same block anchors, byte for byte. Lossless in, review-first out.

01 Grant You open one section to one agent, with limits and an expiry.
02 Read It reads the same files you do, anchored block by block.
03 Propose Its only write verb. Changes land as inert ledger entries.
04 Review Staged notes and diffs wait in your tray, hash-checked.
05 Accept Or reject, or revert later. The pen stays in your hand.

Ten ways to read.

The read-only server. Everything an agent needs to be genuinely useful, nothing that lets it touch your files.

list_documents every note in scope, with ids
read_document one note, byte-exact
read_block one block by anchor id
search_blocks full-text across the scope
list_references links and backlinks
get_transactions the proposal ledger
validate_vault format health check
list_canvases every canvas in scope
read_canvas a canvas as data, not pixels
describe_canvas the canvas, narrated

One way to write.

The proposal server adds exactly one verb: propose_transaction. Every proposal is an inert JSON ledger entry until you decide. This one built the course on the homepage.

{
  "ori_notebook_format": "ori-notebook-agent-transaction-v0",
  "goal": "Make a full thermodynamics course for engineering students",
  "client": { "mcp_client_id": "ori_app", "agent_id": "chatgpt" },
  "intent": { "kind": "create_document", "risk_label": "low" },
  "proposed_patch": {
    "operations": [{
      "op": "create_document",
      "target": { "path": "Courses/Thermodynamics 01 (...).md" },
      "document": { "markdown": "# Thermodynamics 01 (...)" }
    }]
  },
  "user_decision": "pending"
}

What that looks like.

A frontier model drafted a ten-note course into the review tray. One click accepted it; the notes landed rendered.

Agents propose. You review. Nothing writes to disk until you accept.

See the beta status Back to the notebook