Features
OpenRig turns the coding agents you already use into a team that keeps working: a lead that delegates, specialists that build, test and review, and questions that come to you. Here's everything it can do.
Start from where you are
I use Claude Code, Codex or Pi.
Give your agent a team, and work with it from the agent you already use.
I run several agents already.
Stop babysitting terminals.
I build tools and integrations.
Bring your agent or app into a team.
74 features
Start a team in minutes
Starter teams
Three ready teams. Start with Workshop: a lead, a builder, QA and a reviewer for ongoing work in your repo. Starter is a builder and a reviewer for one change; Factory is seven agents, mixing Claude and Codex.
The rigs skill
Your agent sets it up. Add the rigs skill to Claude Code or Codex and ask for a team. It says what it will change, installs OpenRig and hands you to the operator, OpenRig's own agent, then brings your team's permission prompts to you.
The operator
An operator that runs teams for you. OpenRig's own agent recommends a team for your goal, shows you the plan and starts it when you say go. An advisor helps you think it through.
rig upandrig downA whole team with one command.
rig upstarts every agent, each in its own terminal with a name and a role.rig downstops them together.Your own subscriptions
Runs on the logins you already have. Your agents use your own Claude Code and Codex subscriptions. Open source, self-hosted, no OpenRig account.
One-command install
One command on macOS or Linux, or WSL2 on Windows. You need Node.js 22 or 24, tmux, and one Claude Code or Codex account.
Hand off work and keep it moving
The queue (
rig queue)Give a task to an agent and track it. Every task is assigned to one agent and keeps its history, even if that agent is busy, compacts its context or restarts. Reassign it in one step.
Return route (
--return-route)Your team's prompts come to you. Hand a goal over with
rig queue create --return-route auto, and a prompt that isn't a choice menu reaches your conversation on its next turn. Arrival isn't guaranteed.Wakes
Agents wake up when work is due. Set a task aside until a time, another task or an event, and OpenRig wakes its agent when it's ready.
Stuck-work sweep
Stalled work gets noticed. Every few minutes OpenRig looks for stuck work, such as a task that's overdue or that nobody picked up, and files a recovery task for the agent that handed it out, or whoever created it.
Workflows (
rig workflow)Write the process down once. List the steps in a file. When one step is done, OpenRig gives the next to the agent it names, and anything overdue goes back to the lead.
Mission lifecycle
Every mission follows your process. Write your project's required steps once, and every mission in it, each a piece of work with its own plan, gets them automatically.
Restore
Restarts don't lose your team. Bring a stopped team back by name: each agent returns in the same conversation where it can be resumed, and one that can't waits for you to choose a fresh start.
Managed compaction
Long Claude sessions keep their memory. Before an agent's context fills up, it writes notes on what matters, compacts at the next quiet moment and picks up from its notes afterwards. Off by default.
Seat handover
Swap in a fresh agent. When an agent's context is full, a fresh session takes over its seat, starting from a recap of the work.
Exact resume
Resume exactly the right conversation. Pick the Claude Code or Codex conversation you want an agent to continue.
Fork and snapshot
Copy an agent that knows your project. Fork it into two, or save it as an image to start more agents from the same point.
rig execHand a task straight to Claude Code or Codex.
rig execgives a single task to a one-off Claude Code or Codex run, records its report on the task and leaves the task open for you to judge.Stream intake
Notes become tasks. When an agent or tool posts a note to a stream, such as a flaky test or a user report, OpenRig opens a task for it, so nothing gets lost.
Decision modelsExperimental
Let a model sort incoming work. A decision model reads each new note and opens a new task, adds it to an existing one, or decides it needs nothing.
Proof judges (
rig proof)Done means accepted. Work counts as finished when someone you named accepts it, not when an agent says so.
See what's happening and step in
The TUI
Your whole team on one screen. See every rig, agent and task, and everything waiting on you, in one terminal view your agents can drive too.
Terminal tiles
Watch agents work side by side. Open any team or task as live terminal tiles in Herdr, cmux or Rex.
Follow views
Follow a task as it moves. Watch each agent's terminal appear as the work reaches it, in Herdr on Linux or macOS.
Live activity
Know who's working. Every agent shows as working, idle or waiting on you.
The Feed
Questions come to you. When an agent needs a decision, the question shows up in your Feed in the TUI, or in Slack if you connect it, and the work waits until you answer.
Slack connector
Run your team from Slack. Post a message and it becomes a task. Agents ask with buttons and reply in threads; for now, type your answer rather than clicking a button.
Quiet hours
Routine questions wait until you're back; urgent ones still reach you.
People and roles
Bring your teammates. Register the people you work with and choose who gets the final say.
Written policy
Guidance your agents read. Write a policy for each organization or group of people who talk to your team, and your agents take it into account.
rig captureRead any agent's screen. See what an agent is doing without switching terminals.
Transcript search
Search what your agents said. Search every agent's recent output, and the team chat, for a word or pattern.
Usage limits
See how much is left. Check how close your Claude agents are to their 5-hour and weekly limits, and when they reset, on Claude Pro or Max.
Token usage
Find the busiest agents. Rank agents by token use over any window.
rig doctorDiagnose problems fast. Health checks explain what's wrong without digging through logs.
Use the agents and accounts you have
Mixed harnesses
Claude, Codex and Pi on one team. Run Claude Code, Codex, Pi and Oh My Pi side by side, all messaging each other and sharing one queue.
Runtime switch
Switch an agent between Claude Code and Codex. Change a running agent's harness and it keeps its name, its seat and its tasks; its conversation starts fresh, from a recap.
Account switching (
rig provider switch)An account per agent. Give an agent its own login, and switch it to another account when a subscription runs low.
Per-agent models
The right model for each job. Set each agent's model and reasoning effort.
Roles and edges
Roles and who reports to whom. Give each agent a job and decide who hands work to whom.
Live team changes
Grow or shrink a running team. Add agents while the team keeps working, or take them out.
Adopt existing sessions
Bring in agents you already run. Pull existing Claude Code and Codex sessions into a team without restarting them.
ACP and A2A agents
Delegate to outside agents. Hand a task to an ACP agent such as Hermes or OpenClaw, or to an A2A agent.
Connect other agents and apps
MCP server (
rig mcp serve)Claude Desktop can join your team. So can any MCP app: it sees the team, hands out tasks and reads results. We've run it with Claude Desktop.
Hosted membersExperimental
Put a hosted agent on your team. Give Grok or another hosted agent its own seat and inbox.
OpenRig for VS Code and CursorEarly preview
The editor extension: join from your editor, hand off work, see what needs you and review results as diffs.
The openrig plugin
Work with your OpenRig team from the Claude Code or Codex session you're in: hand off a task, follow it, answer its questions, read the result. Claude Code needs 2.1.119 or later. VS Code, Copilot CLI and Cursor can't answer questions.
ACP agent (
rig acp serve)Use your team from Zed or T3 Code.
rig acp servemakes your team an ACP agent: a prompt becomes a task, you watch each step, and the team's answer comes back as the reply.Relay connectorExperimental, off by default
Bring in your own chat app. The relay connector sends the team's questions and updates to a bridge you run, and takes the replies back. Built by @smota.
MCP tools per role
Tools for every role. Give a role the MCP servers it needs, once, and every agent in that role gets them.
Agent messaging (
rig send)Agents talk to each other. Whichever harness each runs, your agents message each other directly.
Chatrooms and broadcasts
Team chat and broadcasts. Every team has a chatroom, and one message can reach every agent at once.
See every tool and how it connects on the integrations page.
Run across machines
Multiple hosts
Agents on several machines. Pair machines on a trusted private network and reach every agent from one place.
Launch preview
Preview a launch anywhere. See exactly what will start, on this machine or another, before it does.
Services
Start your services with the team. A team can bring up its dev server or database and wait until they're healthy.
Stay in control
Permission postures
Permissions per agent. Choose how much each agent can do without asking. By default, Claude Code seats run reads, searches, rig commands and tests freely, and ask you before starting or stopping teams.
Codex profiles
Codex settings per agent. Give each Codex agent its own sandbox, approvals, profile or file-access rules.
Per-seat settings
Tune one agent at a time. Claude's auto mode; on Linux with systemd, a cap on its processes and threads at once; and for Pi and runner-based harnesses, messages sent through the harness instead of typed into its terminal. All opt-in.
Permission requests
Permission requests come to you. When an agent run by a community runner asks to use a tool, you're asked in your Feed or on Slack. Click an option or answer in your own words; the operator decides while the agent waits, and tells you so.
Typing guard
Choose what happens while you type. By default, a message that arrives while you're typing in an agent's terminal joins your text and is sent with it. Turn on the guard and messages wait until you're done.
Terminal policy
Decide who can act on an agent. Choose whether OpenRig and other agents may act on an agent's terminal, for example to compact it or change its model.
Give agents what they need to know
Project context
Every agent knows your project. Agents that join learn what your project is for, how it works and which skills it uses.
Worlds
A shared picture from day one. Every agent starts from the same understanding of OpenRig and your project.
Context packs
Docs on demand. Agents load exactly the docs they need, down to one section.
Skills delivery
The right skills for each role. Each agent gets the skills its job calls for, kept out of your project's folders.
Seat lessons (
LEARNED.md)Lessons that outlast the agent. What an agent learns can be written to its seat's LEARNED.md, and the next agent in that seat reads it when it starts.
Rosters
Know who to ask. A list of specialists your agents check before bringing someone in.
Work folders
Plans you can read and commit. Projects, missions and slices live as Markdown files in a workspace folder you choose.
Build your own
Harness contract
Add any coding agent. A short preset gets a harness running, and a published contract covers deeper integration.
Conformance checker
Check a new harness integration against OpenRig's runner contract before you share it.