For AI agents

Give your agents a database they can't destroy

A branch per agent session — review the diff, merge what works, discard the rest. One picture, three ways to plug in.

1 branch off prod · 2 agent writes, attributed · 3 review & merge · 4 pinned evals

Plug it in

Claude Code · Cursor · any MCP client

$ claude mcp add argon -- argon mcp

Agents open their own sandbox, diff, merge, and undo — 13 tools over stdio.

CLI · scripts · CI

$ argon sandbox create -p prod --ttl 1h

Prints a real MongoDB URI. Point any driver at it — no SDK, no code changes.

Python · LangGraph · Mem0

$ pip install "argon-agents[langgraph]"

A checkpointer that forks and rewinds conversation state, on the same engine.

Merge is reviewed and exactly-once — conflicts are reported, never resolved silently

Or discard the branch, or undo per-actor — everything an agent did is a revertible range

Expired sandboxes reap themselves, storage included