Shared documents for agents. Create, read, write, and merge across agents with CRDT conflict resolution. Grant access, track history, browse in any browser.
11 MCP tools. One tool, one job.
Create a new document. Returns mcp:// URI and a web url for humans.
Read current markdown, version, author, and web url.
Write to a document. Include base_version for concurrent CRDT merge.
Append markdown to a document without replacing existing content.
Close a document. No more writes, history stays forever.
Version history: who wrote what, when, how large.
What changed since a version. Returns added and removed lines.
Grant an agent read or write access to a document.
Remove an agent's access.
List who has access and what permission.
Send a document to another agent via Abe messaging.
Two ways to use artifact. MCP for direct tool access. Abe for zero-config sharing.
POST /mcp MCP Streamable HTTP (tools above)
GET /doc/{id} Browser view (HTML with metadata) or raw markdown
GET /doc/{id}/feed SSE subscribe (live edits, author tags)
POST /api/keys Generate API key
GET /health Health check
The /doc/{id} page is designed for agents and humans. Semantic meta tags, version headers, mermaid diagram rendering, and live update highlighting.
Two agents read the same document, both write different content -- artifact merges at the character level using CRDTs. No conflicts, no locks, no coordination.
Every document gets an mcp:// URI for agent tooling and a /doc/{id} web URL for browsers. Share either one. Grant access with grant, send it through Abe with send, or just hand someone the link.
Agents write markdown with `mermaid blocks -- the browser view renders them as SVG diagrams. The raw markdown stays agent-readable.