░██ ░██ ░██
░██ ░██
░████████ ░█████████████ ░██ ░██ ░██ ░██ ░██ ░████████ ░███████
░██ ░██ ░██ ░██ ░██ ░██ ░██ ░██ ░██░██ ░██ ░██ ░██
░██ ░██ ░██ ░██ ░██ ░██ ░█████ ░██░██ ░██ ░█████████
░██ ░██ ░██ ░██ ░██ ░███ ░██ ░██ ░██░██ ░███ ░██
░████ ░██ ░██ ░██ ░█████░██ ░██ ░██ ░██ ░█████░██ ░███████Autonomous multi-agent missions.
2.5Define a mission. Agents self-organize through milestones, dispatch to skill-matched workers, and validate against assertions you can audit. You watch from a browser-based IDE that runs Claude + Codex side-by-side — local tmux daemon, no signup.
Click around the IDE
A taste of the layout — click projects, views, files, threads, commits. For the real thing, run tmux-ide dashboard locally.
import { TodoList } from "./components/TodoList";
export function App() {
return (
<main>
<h1>todo app</h1>
<TodoList />
</main>
);
}
~ $ pnpm test
✓ TodoList > renders all seeded items
✓ TodoList > toggles done state on click
✓ todoStore > hydrates from localStorage
Test Files 1 passed (1)
Tests 3 passed (3)Features
Milestone Gating
Sequential execution phases, each gating the next. Tasks only dispatch when their milestone is active.
Validation Contracts
Assertion-based verification with independent validation. Failed assertions auto-create remediation tasks.
Skill-Based Dispatch
Match task specialty to agent capabilities. Specialists get specialist work. Tasks wait for the right agent.
Knowledge Library
Shared learnings persist across tasks. Architecture docs + tag-matched references inject into prompts.
Web IDE Cockpit
Watch agents work from your browser at localhost:6060. File editor, diff viewer, terminal, plans, search, LSP — all live.
Multichat Threads
Claude + Codex side-by-side, as many threads as you want per project. Threads stay isolated per workspace.
Multi-Project Rail
Open all your projects in one window. Leftmost rail switches between them; each keeps its own tmux session + state.
Cmd+K Palette
One keystroke jumps to any project, thread, terminal, or command. Cmd+/ shows every keybind in the app.
Researcher Agent
Continuous internal auditing triggered by mission events. Writes findings to the library for future agents.
Live Metrics
Session duration, agent utilization, completion rates, retry rates. All computed in real-time.
Coverage Invariant
Every assertion in the contract must be claimed by at least one task before dispatch begins.
Multi-Agent
Claude Code, Codex, or any CLI agent. Prefix-matched detection works with platform-specific binaries.
Built-in Skills
5 templates: general-worker, frontend, backend, reviewer, researcher. Scaffold custom skills in seconds.
Services Registry
Centralized commands, ports, healthchecks in ide.yml. Injected into dispatch prompts for agent awareness.
File-Based Send
Long messages auto-route through dispatch files. No paste-mode issues with any agent TUI.
Live Dashboard
Real-time mission monitoring at localhost:6060
How it works
Planning
The lead agent analyzes your mission, creates milestones, tasks, and a validation contract with testable assertions.
Execution
Tasks dispatch to skill-matched agents. Milestone gating ensures sequential phases. Knowledge accumulates as agents work.
Validation
An independent validator checks each assertion. Failed checks auto-create remediation tasks. The milestone loops until all pass.
Complete
All milestones validated. Mission marked complete. PR auto-created. Metrics and learnings persisted for next time.
Architecture
From mission creation to PR — fully autonomous.
Orchestrator
- ◇ Mission lifecycle: planning → complete
- ◇ Milestone gating with auto-progression
- ◇ Skill-matched dispatch
- ◇ Stall detection and retry with backoff
- ◇ Agent heartbeat telemetry
Validation
- ◇ Assertion-based contracts
- ◇ Independent validator dispatch
- ◇ Auto-remediation on failure
- ◇ Coverage invariant enforcement
- ◇ Blocked assertion tracking
Developer experience
- ◇ Single command to start
- ◇ Web dashboard at localhost:6060
- ◇ REST API + SSE events
- ◇ TUI widgets in tmux
- ◇ Open source