Dispatch app icon
✺ developer preview

Dispatch

Air-traffic-control for AI agents. Watch the radar, trace methodology, intercept mid-flight, broadcast new protocols across the fleet. Don't just run agents — dispatch them.

macOS 14+ Apple Silicon Local-only Open source v1.4.17
Download for Mac (DMG, 46 MB) Download source (3.8 MB) View on GitHub
First launch — Gatekeeper warning. Dispatch is not yet notarized, so macOS shows "Apple could not verify Dispatch is free of malware" and offers only Move to Trash / Done. To allow it:
  1. Click Done on the warning (do not Move to Trash).
  2. Open System Settings → Privacy & Security, scroll to the bottom — you'll see "Dispatch was blocked…" with an Open Anyway button.
  3. Click Open Anyway, authenticate, then re-launch Dispatch.
Or, from a terminal: xattr -dr com.apple.quarantine /Applications/Dispatch.app. Notarized builds land in a future release.
✺ what it does
live monitoring
The Radar

Glanceable home view — digest rail, team floor of agent cards, live ticker. See every workstream at once.

methodology tracking
The Trace

Forensic trail of why — structured decision events (considered, choice, rationale, confidence) as a timeline, not a transcript.

mid-flight intervention
Intercept

Three modes — nudge (advisory), redirect (hard course-correct), rollback (rewind to a decision and replay with a hint).

skill broadcast
The Protocol

Agent proposes a pattern via propose_skill; you promote it to the team handbook; every workstream's next session adopts it.

persistent memory
The Dossier

Per-workstream Markdown the agent maintains across sessions — identity that survives claude restarts.

summaries on demand
Updates

LLM-generated weekly / monthly digests. Audience-tuned. Claude API or local Ollama. Scheduled or one-shot.

✺ install (60 seconds)
  1. Get the source. Either clone the repo or unzip the download:
    # option a — git git clone https://github.com/S-KSM/Manager.git ~/Code/Manager # option b — download curl -L https://ai-may-i.com/dispatch/dispatch-source.zip -o /tmp/dispatch.zip unzip /tmp/dispatch.zip -d ~/Code && mv ~/Code/dispatch ~/Code/Manager
  2. Run the installer. Idempotent and confirms each step. Builds the daemon, builds the macOS app, copies it to /Applications, wires hooks into ~/.claude/settings.json, registers the launchd agent, migrates legacy v1.1.x state.
    bash ~/Code/Manager/bin/install.sh
    Dry-run prereq check first: bash bin/install.sh --check-only. Skip the app build: --skip-app. Auto-approve: --yes.
  3. (Optional) Enable Updates. For LLM-written weekly / monthly summaries:
    export ANTHROPIC_API_KEY=sk-ant-... # or run a local model: brew install ollama && ollama serve && ollama pull qwen3:8b
  4. Use Claude in any project. Dispatch derives the workstream from the git-root basename:
    cd ~/Code/some-project claude
    A card for some-project appears on the Radar within seconds. Decision events stream into the Trace live.
Prereqs: macOS 14+ on Apple Silicon, Xcode command-line tools, Node.js 18+, and the claude CLI. The installer never sudoes — if a step needs elevation it prints a clear instruction.
Uninstall: bash ~/Code/Manager/bin/uninstall.sh removes the launchd agent, hooks, app, and state under ~/.claude/dispatch/.
✺ where state lives

Everything local-only. The daemon binds to localhost:9876. No auth, no remote access in v1 (both arrive in v1.5 with the iOS client).

~/.claude/dispatch/ ├── db.sqlite # workstream registry, intercepts, skill proposals, reports ├── events/<id>.jsonl # append-only event log per workstream ├── memory/<id>.md # per-workstream Dossier (agent-curated Markdown) ├── handbook.md # team-wide promoted Protocols └── scheduler.json # weekly / monthly report cron config
✺ autonomous mode (v1.4, optional)

By default Dispatch is purely observational — you launch claude, Dispatch watches. With a WORKFLOW.md it also drives: an orchestrator polls a tracker (Linear or a mock JSON file), claims tickets, and spawns claude per turn inside per-issue workspaces. Same hooks, same event store, same Radar / Trace / Intercept.

export LINEAR_TOKEN=lin_api_... dispatch start --workflow ~/dispatch/WORKFLOW.md
✺ why this exists

Today's tooling pushes you into one of two modes: live single-agent supervision (Cursor, Claude Code in a terminal) — high fidelity, but only one agent at a time. Or outcome-based ticket tracking (Linear, Jira) — scales to many, but you only see results, never how the agent got there.

Jira tells you what happened yesterday. Claude Code tells you what one agent is doing right now. Dispatch gives you the high-fidelity oversight to manage an entire department of agents in real-time.