Mainplane: An Open-Source 'Agent Shed' That Keeps AI Agents Working Across Your Own Computers

Mainplane, an open-source project published at github.com/mainplane-ai/mainplane, is a self-hosted control plane that turns any computers a user owns into an always-on "agent shed" so AI agents keep working after a laptop is closed. Mainplane orchestrates agent sessions across worker machines over a built-in secure mesh network, supports any provider's model or an existing Claude Code or ChatGPT Codex subscription, and runs on macOS, Windows and Linux, with a Linux node required for shared file hosting across machines. The project claims a concurrent remote filesystem that lets multiple agents share storage without de-sync and native computer use that drives real browsers and applications rather than a sandbox.

Mainplane: An Open-Source 'Agent Shed' That Keeps AI Agents Working Across Your Own Computers A thread worth a quick look: Mainplane https://github.com/mainplane-ai/mainplane , billed as “the open-source agent ecosystem.” The one-line version that makes it click: it turns any computers you own into an always-on local AI shed — agents keep working when your laptop is closed. If you’ve felt the friction of agentic tools that die the moment you shut the lid, or that trap you on one machine with one provider, this is aimed squarely at that. What It Actually Is Mainplane is a control plane a “harness” for distributed AI agents. It orchestrates agent sessions, manages worker machines the computers your agents control , establishes a secure mesh network between them, and gives you both a terminal UI and other ways to message your agents. Install is a single quickstart command on any always-on machine: Linux, macOS curl -fsSL https://mainplane.ai/install | sh -s -- server quickstart The quickstart walks you through three primitives: a provider the model , workers computers agents control , and connectors ways to message your agents . The Pieces That Stand Out - Bring your own model — or subscription. Mainplane can use any model from any provider Anthropic, OpenAI, Google, … , or the regular limits of a Claude Code / ChatGPT Codex subscription , or a plain API key. That subscription-reuse angle is increasingly the thing people actually want. - Agents + computers, with real security controls. It links agents to machines with fine-grained permissions — no Tailscale needed . The networking is built in. - A concurrent remote filesystem. Multiple agents can use a shared filesystem at once without risk of de-sync — a genuinely hard problem it claims to handle. - Native computer use. It drives your real browsers and applications directly, not a sandboxed facsimile. - Context management. Agents have a restricted toolset and manage their own context , retrieving key info from other sessions rather than drowning in one giant window. - Runs anywhere, even one box. Server runs on macOS, Windows, and Linux; everything can live on a single device shared file hosting across machines wants a Linux node . Why It’s Interesting The center of gravity in agent tooling is shifting from “one agent in one chat window” to persistent, multi-machine orchestration — agents that run long tasks, survive your laptop sleeping, and coordinate across the computers you already own. Mainplane is an explicitly self-hosted, open-source take on that: your machines, your models or your existing subscription , your data not shared by default, opt-in only . It rhymes with other things in this space — reusing a Claude Code / Codex subscription across sessions https://themenonlab.blog/blog/cindy-open-source-agent-reuse-claude-code-codex-subscription is a pattern we’ve seen elsewhere — but Mainplane’s distinguishing bet is the mesh + shared filesystem + native computer use combo that lets a fleet of agents work like a little always-on datacenter in your closet. Worth a look if you’re building agent workflows that outgrow a single machine: github.com/mainplane-ai/mainplane https://github.com/mainplane-ai/mainplane .