The system of record for AI-generated work
Your AI makes great work. Cairn keeps it.
Everything your AI makes: findable, versioned, and ready for its next session. Not buried in chat scrollback.
The problem
AI chat tools are brilliant at making documents and terrible at keeping them.
Buried in scrollback. Finding last month’s brief means scrolling a dead conversation, or asking the model to make it again, slightly differently.
Five versions, none canonical. “Final” and “draft” look identical when they live in five different chats.
Every session starts from zero. You re-paste context as noisy blobs, and the value of yesterday’s work quietly evaporates.
What Cairn does
Four things, done properly.
Search
Everything findable
Full-text and faceted search across every asset your AI has ever made, filtered by type, model, status, project, date. Not by scrolling last month's conversations.
search_assets("pricing") → the approved pricing table, not a chat dump
Versions
One canonical current
One document, an immutable version history, visual diffs. “Final” stops being a guess spread across five conversations.
v1 → v3 · +12 −7 lines · current: v3
Provenance
Every file answers for itself
Which chat produced this. Which model, which prompt, derived from what. On every asset, automatically, with an append-only audit trail.
source: claude.ai/chat/7f3d… · claude-fable-5
Native AI · MCP
Built for the next session
Agents read curated, approved context straight from Cairn and write their output straight back, in Claude, Claude Code, Cursor, anything that speaks MCP.
write_asset(…) → lands in your review queue, not your downloads
Sharing
A browsable library with stable links per asset and per version. Export the whole corpus as portable JSON.
Duplicate guard
“This already exists”: Cairn versions instead of duplicating when an agent re-makes the same file.
Analytics
Retrieval heatmaps, duplicate pressure, what your agents actually reuse.
Audit
Append-only log of every action, human or agent. Governance inherited from legal-grade document management.
The loop
Agents draft. You approve. The next session already knows.
Cairn sits between your agents and you: machine-made work lands as drafts, humans promote what’s good, and every future session reads from the approved tier instead of starting over.
Native MCP
Your agents already know how to use it.
One registration and every MCP client on your machine can search, read, and write your corpus, with provenance captured on the way in.
# register once
claude mcp add --scope user cairn -- node cairn/packages/mcp/dist/index.js
# then, in any session
> search_assets("onboarding email")
> get_asset("a_m5j1q9x3z") # approved v2, not a stale pasteThe tool surface
- search_assets
- get_asset
- write_asset
- new_version
- diff_versions
- set_status
- find_similar
- list_collections
- set_session_context
Works with Claude Desktop, Claude Code, Cursor, and anything else that speaks MCP.
No signup
Use it right now.
The full product experience, seeded with a real corpus, running in your browser. Search it, diff it, approve things, then watch an agent do the same.
Open the sandbox