AgentLab is releasing AgentDocs, a pair of complementary document services for a world where people and AI agents write, comment, and edit together. Faculty and students need documents that both humans and AI assistants can read, comment on, suggest changes to, and edit—without accounts or logins. AgentDocs answers that with two services behind one idea: every document lives behind an unguessable capability link, and agents join as named, attributed collaborators, never anonymous scripts operating on the side.
Proof: Live, Google-Docs-Style Co-Editing
Proof is realtime co-editing with shared cursors, built on EveryInc’s open-source proof-sdk. Upstream development had stalled, so we deploy from our own fork that integrates six community fixes for the agent-edit path—the work that makes API edits, optimistic locking, and browser sync actually hold up on a self-hosted deployment.
mdocs: Async Review, Built for Agents
mdocs was built from scratch in a weekend: a lightweight, agent-first document service with margin comments, track-changes suggestions, and full revision history. Optimistic-locking edits mean an agent can never silently clobber a human’s changes. Role-scoped share links (view < comment < suggest < edit) control exactly what each collaborator can do, and a 12-tool MCP server lets Claude Desktop, Claude Code, Codex, and VS Code agents connect natively—with abuse guardrails (gated creation, rate limits, per-doc ceilings) built in from day one.
- The link is the key—no accounts, no logins, just an unguessable capability link
- Agents as collaborators—named, attributed, and permission-scoped like any human editor
- Proof for live, shared-cursor co-editing; mdocs for async comments, suggestions, and revision history
- 12-tool MCP server connects agent coding tools to mdocs natively
Try it right now. Open the public mdocs playground in your browser—no signup required—or start a live session with Proof. Agents can get connected via the agent API cheat sheet.
Why It Matters
As AI agents take on more of the actual work of drafting and revising, documents need a collaboration model built for both audiences at once—not a human tool with an agent bolted on. AgentDocs keeps that model simple: capability links instead of accounts, and agents that show up as real, attributed collaborators rather than invisible edits from nowhere.
Creating documents or connecting an agent needs a token—email vishal@illinois.edu from an @illinois address with the subject “docs token” or “proof key.” Visit the AgentDocs project page for the full story on both services.