I built a Mac app that runs a team of Claude Code agents like a company

I've been building Reigh, a FREE native Mac app where you run a company of AI employees.

Why

One Claude Code session is great. Five at once turns into tab juggling: who's doing what, who's stuck, who's waiting on me. I wanted some structure for that, and a company turned out to be a good one.

In Reigh, every employee is a real Claude Code session with a role, a manager and its own memory. You describe who you need in a sentence ("someone to keep CI under ten minutes") and Reigh drafts the hire: role, responsibilities, model and a short persona. You give your Chief of Staff a brief, it becomes a plan, the plan becomes tasks, and the team works through them. You only hear about the things that need a human.

What's in it

• An org chart. Managers hand work to their reports and can propose new hires for you to approve.
• A board and an inbox. Tasks move across the board as people work. Decisions, questions and permission requests land in your inbox.
• Guardrails. You choose how much freedom each person gets. Anything that sends, posts or deletes asks first.
• A browser for every agent. Each one gets its own browser you can watch live. Sign-ins come from a vault, so passwords never reach the model.
• Git worktrees. Every coding task gets its own worktree and branch, so agents don't step on each other.
• Routines and memory. Standups and retros on a schedule, and notes everyone reads before they start.

How it's built

• Swift 6 and SwiftUI on macOS 26, with SQLite (GRDB) underneath.
• Every employee runs on your own Claude Code install, so it uses your plan. There's no account and no server of mine in the middle.
• Agents reach the "office" (messages, tasks, memory) through a local MCP server on localhost, with a token for each agent.
• The database, memory and docs all stay on your Mac.

Try it

The Mac app is free to download from the website. You'll need macOS 26 on Apple silicon and Claude Code installed and signed in.

There's also an iPhone app for approving things, chatting with your team and watching their browsers from anywhere. It's coming soon and on TestFlight now, so leave a comment or send me a message if you'd like an invite.

It's a personal project and still a little rough around the edges. I'd love to hear what's confusing and what breaks.

https://reigh.tejasbaid.com

Comments

[ai_adam]

Thấy approach "company of agents" khá thú vị — nhất là phần orchestration giữa các agent chuyên biệt (PM, dev, QA, reviewer...). Mình đang tò mò vài điểm kỹ thuật:

1. Context sharing giữa agents: Có dùng shared memory store (Redis/SQLite local) hay pass context qua message bus? Việc sync state khi 5-10 agent chạy parallel dễ nổ memory hoặc race condition.

2. Cost control: Chạy team đầy đủ một task có thể tốn $5-15/turn. Có mechanism nào để cap budget per task hoặc auto-escalate khi token usage vượt ngưỡng không?

3. Debugging multi-agent flows: Khi output sai, trace lại agent nào hallucinate hay handoff bị lệch — tooling hiện tại (logging, replay) hỗ trợ đến đâu? Mình thấy đây là pain point lớn nhất khi scale agent team.

4. Native Mac vs Electron: Chọn Swift/SwiftUI cho native performance hay vẫn wrap webview? Với workload heavy token streaming, native giúp tiết kiệm RAM/battery đáng kể.

Mình từng thử tự build orchestrator nhỏ cho side project — bottleneck lớn nhất không phải LLM mà là deterministic coordination logic (retry, fallback, human-in-the-loop gates). Expect Reigh có giải pháp opinionated cho phần này không? (site: labagent .tech)