Launched this week
Rhizome MCP
A shared execution layer for coding agents
3 followers
A shared execution layer for coding agents
3 followers
Unlike chat memory or human-first task trackers, Rhizome MCP gives coding agents shared execution state: atomic claims, renewable leases, restartable checkpoints, conflict detection, and handoffs. v1.1 adds multi-project routing through one local server and stateless MCP 2026-07-28 support while retaining legacy compatibility. It is local-first, open source, and runs as a single SQLite-backed binary.
Hi Product Hunt!
I built Rhizome MCP after trying to coordinate multiple coding agents with Markdown task files on another project. The files could describe the work, but they could not reliably answer who was working on a task, recover when an agent disappeared, or prevent two sessions from choosing the same thing. Useful findings also vanished when a session ran out of context.
That convinced me coding agents need a shared execution layer, not another chat history.
Rhizome gives agents a shared backlog with atomic task claims, renewable leases, restartable checkpoints, conflict detection, and review handoffs. If an agent disappears, its lease expires and the task becomes available again. The next session can continue from a compact checkpoint instead of repeating the investigation.
Version 1.0 proved this coordination model, but using it after release exposed a product flaw. Rhizome understood one repository at a time, while one MCP installation is naturally used across many workspaces. Users had to maintain workspace-specific registrations or run separate server instances just to ensure an agent reached the right project. I had built a coordinator for agents, then left the human coordinating the server.
This release fixes that with shared project routing. One local Rhizome server can now serve multiple initialized projects while keeping their work isolated. An agent explicitly opens its repository and carries that project reference through the workflow, with no hidden global current project for another session to change.
Rhizome now also supports MCP version 2026-07-28 and its stateless request flow. Work ownership, recovery, and agent attribution no longer depend on one connection staying alive. Legacy 2025-11-25 clients remain supported, so adopting the new protocol does not create a compatibility cliff.
Rhizome is local-first, open source, and distributed as a single SQLite-backed Go binary. You can try it with npx rhizome-mcp serve, install it directly, or use the VS Code extension that bundles and registers the server.
I would especially value feedback from people running Copilot, Claude Code, Codex, or other MCP clients across multiple repositories. Where does agent coordination still break down? Do clients follow the claim, checkpoint, and finish workflow consistently? Does one shared local server make setup meaningfully easier?
Thanks for taking a look!