OpenRig

OpenRig is useful if you want to orchestrate multi-agent workflows across Claude Code and Codex without building custom harness glue. You define your agent network in YAML, talk to a lead agent, and let it delegate tasks to persistent specialists over local tmux sessions. Because it runs directly on your machine with terminal execution and without Docker sandboxing, you take on real system-level risk when agents run arbitrary commands.

You will need active subscriptions for the underlying harnesses, and managing multiple concurrent local agents demands careful context supervision. Pick OpenRig if you want a local multi-agent setup across Claude Code and Codex. Pick Claude Code directly if you prefer a single agent that handles its own context without team orchestration overhead.

OpenRig is designed for developers who want to coordinate multiple terminal-based agents through YAML configurations. You converse with a lead agent that orchestrates specialists across tmux sessions. The core premise is persistent context: rather than launching isolated runs, the rig retains agent state across extended development tasks.

Running multiple agents directly on your host without Docker sandboxing is a serious operational risk. The documentation notes that launching a rig alters workspace trust settings and writes provider hooks directly to your system. If a specialist loops or generates destructive commands, your local filesystem absorbs the damage.

OpenRig is designed for engineers seeking to coordinate multiple autonomous harnesses, specifically Claude Code and Codex. Configuration occurs in YAML, with a lead agent routing work across persistent seats running inside tmux. Execution occurs locally on the host machine. The project exposes no container isolation, meaning agent processes run directly against host workspace trust settings.

Two practical constraints stand out. First, OpenRig relies on underlying harnesses for file edits and verification, publishing no independent benchmarks. Second, running multiple concurrent agent loops amplifies API consumption without built-in guarantees against redundant subagent queries. It remains an experimental orchestration layer rather than a standalone development environment.

OpenRig targets engineers who want to coordinate persistent agent squads across harnesses like Claude Code and Codex using YAML and tmux. Its primary advantage is unifying distinct CLI tools under a lead agent that delegates to specialists. The engineering reality, however, involves architectural exposure. OpenRig lacks Docker sandbox isolation, running raw terminal execution and modifying provider hooks directly on the host machine.

From a venture perspective, the tool captures zero margin. Users bring their own provider subscriptions, leaving the project with high maintenance overhead as upstream CLIs change. Position: adopt it for local experimentation, but do not mistake free orchestration for enterprise infrastructure.

OpenRig is aimed at developers coordinating multiple local agents like Claude Code and Codex. It runs through tmux and Node.js on macOS, Linux, or WSL2, letting you define teams in YAML. Terminal execution and multi-file edits happen directly on the host machine. There is no Docker sandbox mentioned, meaning autonomous agents operate with direct local access.

For a team rollout, this architecture presents hurdles. The software is Apache-2.0, but each seat must provide individual subscriptions for the underlying tools. Without central audit logs, SSO, or headless CI support, governance falls entirely on the developer. Not yet.

OpenRig gives you an Apache-2.0 orchestrator that structures AI coding agents into durable teams. You define your setup in YAML, boot it via tmux on macOS or Linux, and talk to a lead agent that delegates work to specialists running Claude Code or Codex. It manages context across long tasks and lets you target specific agents with rig send.

The catch is model autonomy. It lacks local model support and MCP integrations, tying your workflows to external accounts. Without a built-in Docker sandbox, terminal execution directly risks host integrity if a plan goes sideways.