OpenRig vs Claude Managed Agents

Why I publish comparisons. This category is new. The vocabulary isn't settled. Nobody has a complete mental model of what's in the space, including me. The comparisons here are an attempt to find the shape of the category: what approaches are being taken, what each approach looks like, what we might learn from each other. Not a bake-off.

Claude Managed Agents started from: what if running Claude agents as coordinated infrastructure could be a managed cloud service that ISVs and product teams integrate through an API? The answer is a hosted runtime. Anthropic operates the containers, the network, the event loop. You define agents as versioned specs with tools, MCP servers, and skills, then drive them through API calls and SSE event streams. Notion, Asana, Sentry, and Rakuten are early customers, companies embedding Claude agents into their own products. The operator of the topology is your application code.

OpenRig started from a different question: what if I could run Claude Code and Codex on my own machine as a coordinated team, attach to any one and talk to it directly, and operate the whole thing like infrastructure I own? The answer is a local control plane. You define a rig in YAML, boot it with one command, and get a topology of persistent tmux sessions you can see in a graph and address by name. The operator is you, at a terminal. The agents are Claude Code and Codex as they ship: interactive TUIs you can attach to, not API endpoints you call.

Neither answer is obviously right. They're different bets about who the operator is. Anthropic bets on the application developer who wants Claude as a managed service inside their product. OpenRig bets on the developer who wants the agents on their own machine, with a CLI in front of them and a graph of the topology open in a browser tab.

Because OpenRig runs on your machine and you operate it directly, a few things fall out:

• You boot a whole rig with one command. rig up product-team gives you orchestration, development, and review pods, seven seats running Claude Code and Codex, ready to take work. The spec library ships proven shapes. You don't design a topology before you can try one.
• You attach to any agent and talk to it. Each agent is a real Claude Code or Codex session in a tmux pane. You read its screen, type into it, watch it work. The exploratory back-and-forth that drives real coding work happens live, not through an event stream.
• You mix runtimes in one rig. Claude Code and Codex coordinate as peers in the same topology. Different models catch different mistakes; OpenRig treats cross-runtime as the default, not the exception.
• You bring in agents you're already running. rig discover finds Claude Code and Codex sessions you launched manually in tmux; rig adopt binds them into a managed rig. Your existing work becomes managed without being rebuilt.
• You save the rig and come back to it. rig down snapshots the topology; rig up <name> restores it. Sessions resume where they were able to; everything else is reported honestly so you know which nodes resumed, which were rebuilt, which started fresh.

None of this is in opposition to Managed Agents. It's just a different shape for a different operator. If you're building a product and need Claude agents as a managed service you call from application code, Managed Agents is the right tool. If you're a developer and you want to run Claude Code and Codex together on your machine and operate them yourself, that's OpenRig. The category is large enough to have both shapes, and probably more.