A document workspace designed around AI from the start

The problem we're solving

When people try to use AI in their writing workflow today, they paste text into a chat window, get a response, and manually copy it back. That's not collaboration — that's copy-paste with extra steps. The document and the AI live in completely separate worlds, and keeping them in sync is entirely the human's job.

Even the tools that try to integrate AI treat it as a second-class feature. There's a special "AI sidebar" or a dedicated "AI mode." The AI can't see what anyone else is doing, can't leave a comment, can't propose an edit that gets reviewed like any other change. It's bolted on.

We think that's the wrong model. If an AI agent is doing useful work on a document, it should show up as a collaborator — visible in the activity log, constrained by the same access rules as everyone else, and accountable for what it changes.

How Insync works

At its core, Insync is a CRDT-backed document — built on Yjs — that keeps every participant perfectly in sync in real time. That includes humans typing in a browser tab and AI agents reading or writing through the Model Context Protocol (MCP).

MCP is an open standard for connecting large language models (LLMs) to tools and data. In Insync, it means any MCP-compatible agent — Claude, GPT-4o, Gemini, or your own internal model — can join a live document session without any custom integration. They call the same tools a human uses: read the document, propose an edit, leave a comment, check who has access.

The document always exists in two forms simultaneously: Rich Text for humans and clean Markdown for agents. Both are live projections of the same underlying CRDT state. When you edit in Rich Text, the agent sees the Markdown update immediately, and vice versa. No sync delay, no conversion step.

Permissions that actually mean something

One of the things we care about most is that the rules you set for humans apply equally to AI agents. There are no bypass paths. If you lock a section so no one can edit it, an agent can't edit it either. If you give a model read-only access to a document, it literally cannot write to it — not even through the MCP API.

Permissions flow from workspace down to document down to individual sections. You can give claude-opus proposal rights on one document while keeping gpt-4o in a comment-only role on another. The same access control UI handles all of it.

Every action is attributed and logged. You can see who wrote what, when, and whether the author was a human or a specific model. If an agent proposes a change, it shows up in the activity log as "claude-opus / Opus 4 proposed an edit" — not as a generic system action. Accountability is built in, not added on.

Edit proposals, not silent rewrites

Agents in Insync cannot silently overwrite your document. Every change an agent wants to make arrives as a proposal — a tracked diff you can review, accept, or reject. You see exactly what changed, why the agent changed it (if it left a comment), and who else has already reviewed the proposal.

This isn't just a safety feature. It's what makes collaboration between humans and agents actually useful. You can ask an agent to rewrite a section and review its work the same way you'd review a colleague's pull request. Accept the parts that are good, reject the parts that aren't, and leave a comment explaining why.

Comments agents can actually read

Inline comments in Insync are structured. When you highlight text and leave a comment, Insync records not just the text of the comment but the exact passage you highlighted, the surrounding paragraph, and which block it sits in.

That means an agent reading the document later can find exactly what you were referring to — even if the surrounding text has changed. Comments aren't floating annotations that drift; they're anchored context that both humans and agents can act on.

Who this is for

Insync is for teams that are already using AI in their work and want something more structured than a chat window. Engineering teams documenting systems. Product teams writing specs that agents help keep up to date. Research teams where multiple models contribute analysis alongside human writers.

It's also for people building agent workflows who want a real document backend — not just file storage, but a live collaborative session their agents can participate in like any other user.

What's next

We're working on making agent onboarding faster — connect an MCP client, point it at a document, and it should be contributing in under a minute. We're also expanding the permission model to cover finer-grained write scopes, so you can give an agent permission to edit only specific sections or content types.

The long-term vision is simple: a workspace where it genuinely doesn't matter whether a collaborator is human or AI. Same interface, same rules, same accountability. Just people and agents getting work done together.