Two months of OpenRig: new features and improvements
By Mike
What's new and what got better in OpenRig from 0.5.15 to 0.7.11, by feature.
OpenRig runs a team of coding agents (Claude Code, Codex, Pi and others) that hand work to each other, keep going while you're away, and bring you the decisions. Between mid-September and 0.7.11 we shipped 30 releases and merged about 2,100 pull requests, hundreds of them from contributors outside the OpenRig team. Together they added more than 40 new features and more than 20 improvements to existing ones. Here's each one, with what it does for you. Entries marked improved existed before September.
Setup
- The
rigsskill: ask Claude Code or Codex to set up OpenRig. It installs and starts OpenRig, says what it's changing, and passes your goal to the operator. - Provider choice (
rig setup --providers): use only Claude, only Codex, or both. One subscription is enough, and OpenRig remembers your choice. - Joining your own session: the
rigsskill offers to make your own Claude Code or Codex session a named teammate, so the lead can hand work to you. - Several teams from one spec: run one spec as several teams with
rig up --name, and change one link on a running team withrig edge add. Thanks to @klueless-io. - The operator (improved): it now asks what you're building, offers three teams with a drawing of each, fits them to the accounts you have, and starts one only when you say yes.
- Starter teams (improved):
- Starter: a Claude builder with a Codex reviewer.
- Workshop: a lead, builder, QA and reviewer, for ongoing work.
- Code-review, research and PM specialists.
Long-running work
- Stream intake (
rig stream): anything an agent or tool reports can become a task on its own. Thanks to @Totopo27 for making that the default. - Decision models: an intake agent can let a model file each new item as a new task, an existing task, or nothing to do. Experimental.
- Escalation edges: when an agent has no lead to ask, its problem goes to whoever you name with an
escalates_tolink. - Restore (improved): after a reboot, starting OpenRig brings the operator back (thanks to @haralez), and one broken agent no longer stops the rest of a team from coming back.
- Conversation resume (improved): after a
/clear, agents now come back to their current conversation, forks get their own, andrig seat historylets you pick exactly which conversation an agent continues. Choosing the exact conversation is @bradvin's work. - Managed compaction (improved): before a Claude agent's context fills up, it now writes a ranked restore map and reads it back afterwards. This survives a daemon restart and can be cancelled. Since 0.7.11 it compacts at the next quiet moment instead of asking again. The restore-map location setting is from @Aschenbath.
- The queue (improved): tasks now close with a written result, so you can see later who did what and why. Since 0.7.11, queue commands also work across machines.
- Seat handover (improved): hand a seat to a fresh agent as many times as you like. The outgoing agent writes a recap first.
- Wake batching (improved): an idle agent with several tasks due gets one combined reminder instead of one per task.
- Watchdogs (improved): retune a watch on a running team (
rig watchdog update), and set context watches by tokens or percentage. Thanks to @AdamWhitehurst and @akushonkamen.
Watching and steering your team
- Follow views: open a task, slice or mission and watch each agent's terminal appear as the work reaches it, in Herdr on Linux and, since 0.7.11, macOS.
- Quiet hours: routine questions wait until you're back; urgent ones still reach you.
- People and roles: register the people you work with, choose who gets the last word, give a Slack shared channel one voice on the team, and give each group a written policy your agents read.
- The TUI (improved): it now opens for you after install with the operator and an advisor. It shows a team as a graph before you start it, and has agent pages with health and recent work. The effort column is from @tianxingjian2406.
- The Feed (improved): agents' questions now land in the TUI's Feed, and you choose whose questions you see.
- Slack connector (improved):
- agents ask with buttons, and typed answers are kept;
- replies thread, and reactions reach the agent;
- each team gets its own channel, and one Slack app covers all your machines;
- messages sent while OpenRig was down arrive when it's back.
- Usage and cost (improved):
rig usage costshows what each agent costs, as its harness reports it.rig ps --nodesshows the model it last reported using, andrig ps --resourcesshows your machine's load. - Terminal providers (improved): Rex joins Herdr and cmux (the Rex provider is @vanducng's). You can open views from inside cmux, and tiny panes read at a readable size.
Agents, models and accounts
rig exec: hand one task straight to Claude Code or Codex. It runs a turn and brings back its report, and accepting the work stays with you.- Runtime switch (
rig seat set-runtime): move an agent between Claude Code and Codex mid-job. It keeps its name, its tasks and its place, and starts a fresh conversation from a handover. Built on @Aschenbath's design, after a request from @reco. - Account switching (
rig provider switch): give each agent its own login, and move it to another account when a subscription runs low. Since 0.7.11 this works for Pi seats too. - Oh My Pi: run Oh My Pi agents alongside Claude Code, Codex and Pi. Oh My Pi support is from @MelonSmasher.
- Community harness presets: ready presets for Cursor's agent, Hermes, OpenClaw, Pi's terminal UI and DeepSeek Harness, and a short YAML file adds your own. The dsh-seat runner is @bpamiri's.
- ACP and A2A delegation: hand a task to any ACP agent with
rig acp run, or to A2A agents with a recipe. - Per-seat settings: set one agent's reasoning effort, working folder or environment without editing the team's spec, and give a Claude member its own agent file and prompt. The working-folder setting finishes work by @Aschenbath, who also built the member setup.
- Mixed teams (improved): the Starter team now pairs a Claude builder with a Codex reviewer out of the box.
Connecting other assistants and apps
- The openrig plugin (new in 0.7.11): work with your team from your own Claude Code or Codex. Hand off a task, follow it, answer the team's questions and read the result without leaving your agent. It also installs in VS Code, Copilot CLI and Cursor, but there it can only hand off work: it can't claim tasks or answer questions.
- ACP agent,
rig acp serve(new in 0.7.11): use your team from Zed or T3 Code. A prompt becomes a task, you see each step, and the team's answer comes back as the reply. The@openrig/acplauncher adds it from the ACP registry. - OpenRig for VS Code and Cursor: the editor extension. Join from your editor, hand off work, see what needs you and review results as diffs. Early preview.
- Hosted members: give Grok or another hosted assistant its own seat and inbox, and it picks up tasks like a teammate. Since 0.7.11, assistants connected through a grant service can also answer the team's questions. Experimental, off by default.
- Relay connector (new in 0.7.11): @smota's connector sends your team's questions and updates to a chat bridge you run, and takes the replies back. Experimental.
- MCP server,
rig mcp serve(improved): Claude Desktop and other MCP apps can now work your team's queue: see who's there, hand out tasks and read results.--httpserves it to another machine, with a separate grant for each app. - Daemon HTTP API (improved): new reads for building your own client, two of them from @dajiaohuang.
See every tool and how it connects on the integrations page.
Multiple machines
- Launch preview (
rig launch --plan --host): see exactly what will start on another machine before it does. - Git over SSH: agents push and pull through your SSH keys.
- Usage across machines: see usage from all your machines in one view. Off by default.
- Hosts (improved): edit and remove hosts from the command line, add them straight from your ssh config, and since 0.7.11 a denied pairing request gives the other machine no access.
- Bundle services (improved): a team can bring up its own Docker Compose services. Off by default.
- Platforms (improved): NixOS works, WSL2's first start is fixed, and OpenRig is lighter on Linux. Since 0.7.11, restarting the daemon's systemd service on Linux leaves your agents running.
Sharing teams
- Login files stay private: shared bundles leave your login files out unless you ask for them.
- Bundles (improved):
- install a team from a GitHub folder link (
rig up <link>), pinned to one commit; - see its models, permissions, files and everything it will write first (
rig bundle inspect); - pick a configuration;
- reinstall safely with backups;
- list your team on openrig.dev/rigs with one pull request.
- install a team from a GitHub folder link (
- Export (improved):
rig exportnow keeps the team's culture and startup settings. Keeping the culture file and rig-level startup is @Totopo27's fix.
Permissions and safety
- Typing guard and draft delivery:
- a message that arrives while you're typing joins your text and goes through (since 0.7.11);
- turn on the typing guard and messages wait until you're done;
- its urgent-only mode lets only urgent messages through.
- The draft-aware mode is @Aschenbath's, re-landed with @Mr-Neutr0n's commits.
- Default allowances: routine commands, project reads and common tests run without a prompt; starting and stopping teams still asks.
- Tool rules per agent: separate Claude allow and deny lists, or Codex sandbox and profile settings, for agents that share a folder. Codex profiles were @bpamiri's request.
- Terminal policy (
rig seat set-terminal-policy): decide how far other agents may step into an agent's terminal. - Permission requests: when a runner agent asks permission, you answer in the Feed or Slack. Since 0.7.12, your answer, in your own words if you like, goes to an agent that reads it with the question and decides while the runner waits.
- Hidden credentials: keys OpenRig passes to agents don't show up in terminal or process listings. @pr0xc3nt4ur1 first raised it.
- Web UI lockdown: the web interface is off unless you turn it on, and it checks every request's origin.
- Permission postures (improved): set each agent's permissions with an audited command, and see what it will run with before it launches.
Building integrations
- Harness contract: add any coding agent. A short preset gets it running, and the published contract covers deeper integration.
- Runner conformance checker: test a runner before you share it, with no daemon, model or network needed. Its
runnerfield follows @bpamiri's proposal. - Telemetry reads (
rig telemetry): read events and history without SQL.
Smaller improvements
- Previews:
rig launch --planfor one agent, andrig downasks before stopping the team you're in. - Clean repositories:
- keep OpenRig's Claude instructions in an untracked
CLAUDE.local.md(from @hvpaiva); - team skills live outside your project (skills by @puniserv);
rig downremoves what OpenRig added.
- keep OpenRig's Claude instructions in an untracked
- Faster command startup: CLI commands start several times faster.
- Terminal input: Shift+Enter makes a newline, Ctrl+G opens your editor, and image paste works in agent terminals.
- Rig names instead of IDs, everywhere. Thanks to @Nikhi00718.
- Upgrade notes for agents: after an upgrade, your agent tells you what changed for you.
Thank you to everyone who contributed: every release in What's new names them. Ready to try it? Start here.