Who
I still run multi-agent days where the bottleneck is the shell context, the task queue, and whether each agent can see my real ~/.zshrc aliases instead of a sandboxed stub.
For operators wiring local agent pools against a Mission Control ship list on their own machine.
What
I shipped HurleyUS Agent OS, public https://github.com/michaelmonetized/agent-os. HEAD f66fffa. 2 commits. package hurleyus-agent-os@1.0.0.
orchestrator.ts is an Ink React terminal UI plus a WebSocketServer on ws://localhost:9999. Agents connect with ?agent=<name>, pull task_dispatch, and stream task_output / complete / fail. State lives in ~/.hurleyus/agent-os/state.json; every session transcript lands in ~/.hurleyus/agent-os/sessions/<sessionId>.log.
agent-client.ts spawns interactive zsh (zsh -i -c …), sources ~/.zshrc, optionally loads ~/.hurleyus/GOALS and ~/.hurleyus/TASKS, then evals the task description. If the cwd is a git repo it auto-commits [agent-name] <taskId> and pushes origin HEAD.
mission-control-tasks.ts seeds seven Phase 1 tasks aimed at hurley-mission-control: Convex backend, web layout, architecture review, message wiring, polling, Vercel deploy, QA smoke. cli.ts load-mission-control / start.sh load that list.
Same tree also has core.ts: four personas (Codex Dev, SR Designer, SR Architect, QA Auditor) on claude-opus-4-6 via Anthropic Messages API, with a dependency-aware priority queue. Honest gaps: @anthropic-ai/sdk is imported but missing from package.json; .env with ANTHROPIC_API_KEY is still tracked after the .gitignore commit (gitignore only covers node_modules/, *.log, .DS_Store). OpenClaw relay to ws://192.168.1.134:18789 is stubbed. agent-runner.sh is a file inbox/outbox fallback, not live WS.
Sibling tools in the same shop: orclawstrator's OpenClaw :3377 gateway, the Go p10k mission-control portfolio TUI, the Convex human|agent hurley-mission-control web plane, and shagent's MCP/OpenRouter loop. Agent OS is the Ink/WS/zsh control plane with a Mission Control Phase 1 seed.
Where
Code: github.com/michaelmonetized/agent-os, public, no LICENSE file (README: internal HurleyUS).
git clone https://github.com/michaelmonetized/agent-os.git
cd agent-os
bun install
# Terminal 1
bun cli.ts run-orchestrator
# Terminal 2+
bun cli.ts spawn-agent codex-dev
bun cli.ts spawn-agent sr-designer
# Load Phase 1
bun cli.ts load-mission-control
bun cli.ts status
Local only: orchestrator ws://localhost:9999. Optional env OPENCLAW_GATEWAY. Runtime state under ~/.hurleyus/agent-os/.
When
Created on GitHub 2026-03-20. First commit 11:13 AM ET same day: full v2 surface (Ink + WS + zsh + CLI + Mission Control seed). Second commit 10:25 PM ET: .gitignore only. GitHub pushed_at 2026-03-21T02:25:13Z. Pack prepared 2026-09-08 ~5:17 PM ET. Draft + assets only.
Why
I needed a shell-native control plane that treats agents like terminals with jobs. Ink for the operator view, WebSocket for fan-out, real zsh so aliases and git muscle memory stay intact, and a Mission Control Phase 1 seed so the queue is not empty on day one. The Anthropic core.ts path is the API-side twin when you want model output without a shell; the gaps (missing SDK dep, tracked .env) are the honest day-one scars.
Engagement Q: How would you unify the Ink/WS/zsh path and the Anthropic core.ts path: one CLI surface, or keep them as two intentional modes?