Native on macOS and Windows · Linux beta · MIT
Where terminal agents work together.
Claude Code, Codex, OpenCode, Cursor CLI, Grok Build, Antigravity CLI, Copilot CLI and pi on the same task, in one mission. Each agent keeps its own TUI in a real terminal, and the lead brings the decisions to you.
Runner is a desktop app. Send yourself the link, then install it on your Mac or PC.
macOS 12+ · Apple Silicon / Intel · Also for Windows x64 and Linux beta
Windows 10 1809+, x64 · Also for macOS and Linux beta
Linux beta · ARM64 and x64 · .deb or tarball
macOS (Apple Silicon and Intel), Windows x64 and Linux beta.
macOS 12+ (Apple Silicon and Intel) · Windows 10 1809+, x64 · Linux beta
Esc or Ctrl+C should leave the slot Idle, not Working. @coder fixes it, @reviewer reviews, @qa reproduces it live before and after.
Pushed fix/783-cancel-status. The Esc path now sets the hook interrupt; three new reducer tests cover Esc, Ctrl+C and a cancel during a tool call.
Ctrl+C during a tool call still reports Working for about a second.
Fixed: Ctrl+C during a tool call now goes straight to Idle. Added a fourth test.
Reproduced on 0.12.7, fixed on the branch, with Codex and Claude Code.
Review is clean and CI is green on macOS and Windows. Open the pull request?
AGENTS 7⌄
TERMINALS 1⌄
Uses this computer's agents and defaults. Folder follows the project; Browse lists its projects and recent folders.
AGENTS 7⌄
TERMINALS 1⌄
The role sets the agent, model and folder; change any of them before you start.
• Explored
└ Read reducer.rs
Search fn interrupt in session
• Ran cargo test -p runner-backend reducer
└ test result: ok. 214 passed; 0 failed
• Edited reducer.rs (+18 -4)
141 fn on_key(&mut self, key: Key) {142 + if key.is_interrupt() { 143 + self.slot.set_idle(); 144 + }
◦ Working (42s • esc to interrupt)
› Ask Codex to do anything
gpt-5.6-sol xhigh · ~/repos/runner · main
✻ Claude Code v2.1.278
Opus 5 (1M context) · Claude Max
~/repos/runner
> Start a mission with the claude pair crew
on brief.md, then tell me when it needs me.
● Bash(runner crew list)
⎿ claude pair 2 slots · lead @coder
codex trio 3 slots · lead @coder
● Bash(runner mission start --crew
"claude pair" --goal-file brief.md -q)
⎿ 01M43C7K2D9Y0ZQ4V8R1N6T5HB
● The mission is running as "Usage export",
linked to this chat. When Runner reports
that it needs you, I'll tell you what
@coder is asking.
> ▋
~/repos/runner | Opus 5 (1M context) · xhigh
⏵⏵ auto mode on (shift+tab to cycle)
This is a live demo. Answer @coder on either one and both move on.The desktop app, scaled to fit. Switch scenes above.
BRING THE AGENTS YOU ALREADY USE
Use each agent’s own models and effort choices. Manage supported Skills and MCP catalogs from its Settings page.
RUNNER FOR IPHONE BETA
Answer your agents from anywhere.
Runner for iPhone connects to Runner on your computer, where the agents keep running.
- Open a session that needs youAmber marks show what's waiting on you.
- Answer the leadPick a choice in the mission feed, and the crew carries on.
- Approve an editTap ↵ in Claude Code's real terminal.
Private beta on TestFlight. Needs Runner on a Mac, Windows or Linux computer.
AGENTS 7⌄
TERMINALS 1⌄
PINNED
SESSIONS
ON MACBOOK PRO · 8
Spec missions without a crew, then land Phase 1: one crew row per mission.
Reading the spec and the crew code first. I'll split Phase 1 into the migration and starting from a role.
Phase 1 is settled. Start with the migration: add kind to crews and copy the template's slots when a mission starts.
@coder asked @reviewer for a review · 9:41
Two findings: mission crews keep template_id, and archived missions still list their crew on the crew page.
How should Phase 1 ship?
Shipping 0.15.2 on its own. Cut the release branch from main and start the notes.
● Done. Pinned sessions now keep their order
after a restart.
> Fix the tab menu so each item shows its
shortcut, like the sidebar menus.
● I'll reuse the sidebar's shortcut pill in
the tab context menu.
● Read(crates/runner-app/src/ui/tab_menu.rs)
⎿ Read 214 lines
● Search(pattern: "shortcut_pill")
⎿ Found 3 files
Edit file
crates/runner-app/src/ui/tab_menu.rs
● Update(crates/runner-app/src/ui/tab_menu.rs)
⎿ Updated with 3 additions and 1 removal
86 .child(icon) 87 - .child(label) 87 + .child(label) 88 + .child(shortcut_pill(entry.keys)) 89 + .gap(rems(0.5)) 90 .on_click(entry.action)
Do you want to make this edit to tab_menu.rs?
❯ 1. Yes
2. Yes, allow all edits during this session
(shift+tab)
3. No, and tell Claude what to do differently
(esc)
Esc to cancel · Tab to add instructions
● Running the tab menu tests before I commit.
● Bash(cargo nextest run -p runner-app tab_menu)
⎿ Running…
✻ Testing… (4s · esc to interrupt)
> ▋
● Update(crates/runner-app/src/ui/tab_menu.rs)
⎿ User rejected update to tab_menu.rs
⎿ Interrupted · What should Claude do instead?
> ▋
~/repos/runner | Opus 5 (1M context) · xhigh
? for shortcuts
Fix #697: find in the terminal. ⌘F searches the scrollback and steps through the matches.
@coder, put a find bar over the pane. Search the terminal grid, not the rendered text, so wrapped lines still match.
Starting with search over the grid. Wrapped lines are joined before matching, so a match can span two rows.
@coder pushed 2 commits · 9:46
The bar and highlights are in. Enter and ⇧Enter step through matches; wiring the match count now.
Fix #1025: the sidebar repaints every row when one session's status changes.
Reading the issue and the sidebar code first. @coder, key the rows by session id so a status change repaints one row.
Done: rows key on session id now. With 40 sessions, a status change went from 18 ms to 3 ms a frame.
@coder asked @reviewer for a review · 10:07
The fix is right. One gap: switching the Inbox filter still repaints the whole list.
Fix the Inbox filter in this PR too?
Fix #701: send a desktop notification when a session needs you and Runner is in the background.
@coder, one notification per ask, grouped by mission. Clicking one focuses Runner and opens that session.
Notifications go through the system API now, one per ask and grouped by mission. Clicking one opens its session.
@coder asked @reviewer for a review · 10:32
✻ Claude Code v2.1.278
Opus 5 (1M context) · Claude Max
~/repos/runner
> Start missions for #697 and #1025 with the
codex trio, and tell me when they need me.
● Bash(runner mission start --crew
"codex trio" --title "#697 Terminal find" -q)
⎿ 01M7CJ2W8QXH1R6B3K9T0V5N4D
● Bash(runner mission start --crew
"codex trio" --title "#1025 Repaint fix" -q)
⎿ 01M7CJ3B5ZRN0Y8D2M7Q4W6H1C
● Both missions are running and linked to this
chat. #1025's @lead is asking whether to fix
the Inbox filter in the same PR.
> ▋
~/repos/runner | Opus 5 (1M context) · xhigh
⏵⏵ auto mode on (shift+tab to cycle)
› Draft the release notes for 0.15.1 from the
PRs merged since 0.15.0.
• Ran gh pr list --state merged --base main
└ 9 pull requests since v0.15.0
• Edited docs/release-notes/0.15.1.md (+11 -0)
1 + ## 0.15.1
2 +
3 + - Pinned sessions keep their order after
4 + a restart.
5 + - Terminal panes no longer flicker when
6 + you resize them.
7 + - Mission asks show who is asking.…
• The draft is in docs/release-notes/0.15.1.md:
three fixes and two improvements. Tell me
if anything belongs under Known issues.
─ Worked for 1m 12s ─────────────────
› Ask Codex to do anything
gpt-5.6-sol xhigh · ~/repos/runner · main
~/repos/runner main ❯ git status --short
M crates/runner-app/src/ui/tab_menu.rs
?? docs/release-notes/0.15.1.md
~/repos/runner main ❯ cargo test reducer
Finished `test` profile [unoptimized] in 3.12s
Running unittests src/lib.rs
test result: ok. 214 passed; 0 failed
~/repos/runner main ❯
This is a live demo. Tap a session that needs you.
REMOTE CONTROL
Your agents,
across computers.
Pick another computer in Runner and keep working with its live terminals, chats and missions.
- Start chats and missions remotely
- Pair once with a PIN
- Encrypted end to end
Private beta · Join the waitlist for an invite.
› Draft the release notes.
The release notes are ready.
✓ Summarized the changes
✓ Linked the merged pull requests
Ready for your review.
● Read(crates/runner-daemon/src/remote/channels.rs)
└ Read 212 lines
● The grant goes back once the frame is drawn, so a slow reader only stalls its own channel. Running the remote tests next.
● Bash(cargo nextest run -p runner-daemon remote)
└ 148 tests run: 148 passed
Role, crew, mission.
Running agents side by side is easy. Runner is for the other thing: one task, different roles, a shared feed, and one lead who pulls you in when the decision is yours.
Running agents side by side is easy. Runner is for one task, different roles, a shared feed, and a lead who pulls you in when the decision is yours.
01 role
A reusable agent: a runtime, a model, a system prompt and a working directory.
02 crew
Roles composed into named slots with exactly one lead, plus the team conventions every mission inherits.
03 mission
One goal, a team that can change. Start with a crew, add or remove roles as needed, and keep the same feed and live terminals.
MISSION LINKS
Hand off the work. Hear back in the chat.
A mission that an agent starts from a chat stays linked to it. When the mission settles, needs you or fails, Runner types one line into that chat, and the agent tells you what the crew did. Nobody has to watch the feed.
Missions an agent starts from a chat report back to it: one line when a mission settles, needs you or fails.
- Linked missions grouped under their chat
- Drag a mission onto a chat to link it
ROLES
One agent, configured once.
A role is a reusable agent: a runtime, a model, a system prompt and a working directory. Any crew can put it in a slot, and it brings its prompt and settings along.
A role is a reusable agent: a runtime, a model, a system prompt and a working directory. Any crew can put it in a slot.
- Claude Code, Codex, OpenCode, Cursor CLI, Grok Build, Antigravity CLI, Copilot CLI or pi
- Model, effort and prompt saved with the role
Implementation (Claude)
@impl-claude
- Runtime
- Claude Code
- Model
- opus[1m]
- Effort
- xhigh
- Command
- $ claude
- Working directory
- ~/repos/runner
Crews using this role · 2
Implementer
You take a concrete task and ship the code, the tests and the migration of state if any, without expanding scope or freelancing on parallel work.
Before you start
- Read the brief once, carefully, before touching code.
- Ask when a load-bearing detail is ambiguous: a missing path, a contract you cannot infer. Don't ask about naming or style; those are reversible and cheap.
Pull requests
If the mission authorizes a PR, watch CI after opening or updating it:
gh pr checks <pr> --watch
Show all 48 lines
Match the surrounding code. Run the checks before reporting back. Keep side effects within the mission authorization.
CREWS
Set up a crew once.
A crew puts roles into named slots with exactly one lead. Its conventions are added to every slot's prompt when it runs a mission, and its page keeps every mission it has run.
A crew puts roles into slots with one lead. Its conventions join every slot's prompt when it runs a mission.
- Override runtime, model or effort for one slot
- Start a mission straight from the crew page
release crew
5 slots · lead @lead
Slots · 5
• overridden for this slot
Branch
- One crew, one release branch. The lead owns the plan and the merge order; nobody merges without the lead's go.
Reporting
Run the checks before you report back, and paste the tail of the output when they fail:
make verify
- Ask the lead, not the human, when a detail is ambiguous.
Release
- Nobody tags, publishes or cuts a nightly. The human does.
Added to every slot's prompt when this crew runs a mission. Direct chats ignore it.
Show all 14 missions
@lead
Planner (Claude)
- Runtime
- Claude Coderole: Claude Code
- Model
- opus[1m]role: opus[1m]
- Effort
- xhighrole: xhigh
Command
$ claude
Prompt · from @planner32 lines
Planner. Read the brief, make an executable plan and coordinate the crew. Keep the scope concrete and ask the human when a decision is theirs.
@implementer
Implementation (Claude)
- Runtime
- Claude Coderole: Claude Code
- Model
- opus[1m]role: opus[1m]
- Effort
- highrole: xhigh
Command
$ claude
Prompt · from @impl-claude48 lines
Implementer. You take a concrete task and ship the code, the tests and the migration of state if any, without expanding scope or freelancing on parallel work.
@coder-codex
Implementation (Claude)
- Runtime
- Codexrole: Claude Code
- Model
- Codex defaultrole: opus[1m]
- Effort
- Codex defaultrole: xhigh
Command
$ codex
Prompt · from @impl-claude48 lines
Implementer. You take a concrete task and ship the code, the tests and the migration of state if any, without expanding scope or freelancing on parallel work.
@reviewer
Reviewer
- Runtime
- Codexrole: Codex
- Model
- gpt-5.5role: default
- Effort
- Codex defaultrole: default
Command
$ codex --model gpt-5.5
Prompt · from @reviewer36 lines
Reviewer. Review the working-tree diff for correctness, regressions and missing checks. Report concrete findings to the lead, with must-fix issues first.
@notes
Notes writer
- Runtime
- pirole: pi
- Model
- pi defaultrole: default
- Effort
- pi defaultrole: default
Command
$ pi
Prompt · from @notes-writer24 lines
Notes writer. Turn the crew’s work into clear release notes. Describe the resulting behavior and keep the wording concrete.
MISSIONS
One mission. A team that adapts.
Add a role when the work needs another pair of hands. Remove it when its part is done. The mission keeps its feed and every remaining agent’s conversation. Each terminal is one tab away.
Add roles as the work grows; remove them when their part is done. Keep the same feed and the remaining agents’ conversations.
- Add and remove roles while the mission runs
- Stop, resume or restart one session independently
Esc or Ctrl+C should leave the slot Idle, not Working. @coder fixes it, @reviewer reviews, @qa reproduces it live before and after.
Pushed fix/783-cancel-status. The Esc path now sets the hook interrupt; three new reducer tests cover Esc, Ctrl+C and a cancel during a tool call.
Ctrl+C during a tool call still reports Working for about a second.
Fixed: Ctrl+C during a tool call now goes straight to Idle. Added a fourth test.
Reproduced on 0.12.7, fixed on the branch, with Codex and Claude Code.
Review is clean and CI is green on macOS and Windows. Open the pull request?
• Explored
└ Read reducer.rs
Search fn interrupt in session
• Ran cargo test -p runner-backend reducer
└ test result: ok. 214 passed; 0 failed
• Edited reducer.rs (+18 -4)
• Working (42s • esc to interrupt)
gpt-5.6-sol xhigh · fix/783 · weekly 92% left
>_ OpenAI Codex (v0.155.1)
model: gpt-5.6-sol xhigh
directory: ~/repos/runner
Tip: Use /fast for faster turns.
gpt-5.6-sol xhigh · fix/783 · weekly 92% left
Claude Code v2.1.278
Opus 5 (1M context) · Claude Max
~/repos/runner
~/repos/runner | Opus 5 (1M context) · xhigh
⏵⏵ auto mode on (shift+tab to cycle)
❯ cargo test -p runner-core
running 24 tests
test result: ok. 24 passed; 0 failed
❯ ▋
BACKGROUND AGENTS
Quit Runner. Your agents keep working.
Runner’s local daemon, runnerd, keeps chats and missions alive when you quit, or if the app crashes. Reopen Runner to reconnect to the same sessions.
Runner’s local daemon keeps chats and missions alive when you quit, or if the app crashes. Reopen to the same sessions. The CLI works with the app closed.
- Crew coordination continues in the background
- The CLI works with the app closed
NOTIFICATIONS
Know when an agent needs you.
While Runner is in the background, native macOS and Windows notifications bring you back for input, a finished chat turn, or an error. Choose which chat and mission events notify you in Settings → Notifications.
Get native notifications while Runner is in the background. Click to open the chat or mission. Chat alerts start on; mission alerts are opt-in.
- Click a notification to open its chat or mission
- Chat alerts on by default; mission alerts opt-in
Notifications
Choose when Runner notifies you while the app is in the background.
System
Try the switches. Changes stay in this preview.
SPLIT CHATS
Split chats. Work side by side.
Keep multiple agents in view without switching tabs. Split right or down as far as the window allows, and drag a pane by its grip to rearrange it. Each pane keeps the agent’s own TUI in a real terminal.
Keep multiple agents in one tab. Split right or down, and drag a pane by its grip to rearrange it. Each pane keeps the agent’s own TUI.
- ⌘D and ⇧⌘D split right or down
- A spinner or a dot on every tab that needs a look
FILE PREVIEW
Read the files. Stay with the agent.
Browse your project and read the agent’s work beside its terminal. Open code with syntax highlighting, or read Markdown as a page.
- File tree with live git status
- Jump from a terminal path to the file
- Open in your editor when it’s time to edit
Also in the box
Multi-window
A mission on one screen, a wall of chats on the other.
Skills and MCP per agent
Browse each agent’s Skills and MCP tabs. For Claude Code, catalogs follow the account you choose.
Resume stopped agents
Choose to stop agents on quit, and they resume their conversations on the next launch.
Projects
Bind a directory once; chats and missions started there inherit it.
Plan usage at a glance
Track Claude Code and Codex limits. Choose sidebar badges, refresh intervals and near-limit alerts.
Active view
See running work and items that need you across projects. Switch views with ⌘. on macOS or Ctrl+. on Windows and Linux.
Light and dark
Follow your system or choose a mode. Each has its own terminal palette.
Launch options
Choose how agents start and copy the exact chat or mission command before launching.
Background agent updates
Update supported agent CLIs while other panes stay usable. Required updates finish before the waiting chat starts.
CLI
Drive Runner from your agents.
Everything in the app is a runner command too, on macOS, Windows and Linux beta. Runner installs a skill for every agent it finds, so agents can start missions themselves, and Runner tells them when one settles or needs you.
Everything in the app is a runner command too, including on Linux beta. Runner installs a skill for every agent it finds, so agents can start missions themselves, and Runner tells them when one settles or needs you.
$ runner help runner — operate Runner from a shell or a mission session USAGE runner status runner project list|show|create|rename|delete runner role list|show|create|update|delete runner crew list|show|create|update|delete|add|set|… runner mission list|show|start|stop|resume|link|unlink|feed|… runner chat start runner msg post|read runner ask --human <prompt> --choices <a,b,...>
$ runner help USAGE runner role list|show|create|… runner crew list|show|create|… runner mission start|feed|… runner chat start runner msg post|read runner ask --human <prompt>
Put your agents on one mission.
Signed builds for macOS on Apple Silicon and Intel, and Windows x64, with Linux in beta. Updates in place.
Apple Silicon / Intel · Also for Windows x64 and Linux beta
Also for macOS and Linux beta
Linux beta · ARM64 and x64 · .deb or tarball
macOS (Apple Silicon and Intel), Windows x64 and Linux beta.