Connect any MCP-compatible agent and it appears in your document like a teammate — with a name, a cursor, and a permission level. No copy-pasting outputs. No waiting for a model to finish. Just work, together.
Last edited just now · —
Every participant — human or AI — edits the same underlying document. Changes are merged automatically using CRDT, so nothing is ever lost or overwritten.
An agent isn't a chatbot you paste results from. It gets a seat at the table: a name in the presence bar, a live cursor, and a role you set.
Every edit, comment, and proposal is logged against a named actor — human or model. Permissions work the same for both. No hidden writes, no anonymous changes.
Insync is built around a live session. Agents and humans join the same document, see each other's cursors, and operate under the same permission rules. There are no wrappers, no middleware, no copy-paste handoffs.
Every access rule applies equally to humans and agents. Lock a section so no model can touch it. Give one agent read access while another can propose edits. There are no special bypass paths — agents work inside the same boundaries as everyone else on the team.
Claude, GPT-4o, Gemini, or any MCP-compatible model — connect it and it joins the live session alongside you. It reads the document, sees your cursor, and writes directly. No generate-then-paste loop. No polling. Just another participant in the room.
Yjs CRDT keeps the document consistent across every tab and every agent session. Two cursors in the same paragraph is fine. Changes from three different models at once is fine. Nothing is ever lost — merges happen automatically, below the surface.
One agent drafts a section, another reviews it, a third proposes a revision. All in the same session, with full attribution. The document is the coordination layer.
Highlight any text and leave a comment. Each thread records the selected text, the surrounding paragraph, and the block it anchors to. An agent reading the document later can find the exact passage without guessing — no fuzzy matching, no context loss.
Every edit, comment, and proposal is logged with a named actor — human or model — and a timestamp. Access is scoped per workspace, document, and section. When something changes, you always know exactly who made it and when.
Switch between rich text and raw Markdown in one click. Both views render the same underlying document — agents always read structured Markdown with no HTML noise or editor artifacts.
No agent can silently rewrite your document. Every change arrives as a proposal with a diff. You review it, accept what you want, and reject the rest. The document stays yours.
Set access at the workspace level and it flows down automatically. Lock a section so no agent can touch it. Give one model write access to a specific document while keeping others read-only. Every rule applies to humans and agents equally — there are no special bypass paths for AI.
Open a document, connect an MCP client, and invite your first agent. It takes about two minutes.