manaflow-ai-cmux-testing-repo-3: The Terminal for the AI Middle Manager: Inside cmux

How a native macOS application uses libghostty to tame the chaos of parallel coding agents.

7 min read • View on GitHub • More from manaflow-ai

A 1920s telephone switchboard operator plugging cables into glowing mechanical brains instead of phone jacks. This illustrates the shift from writing code to orchestrating multiple AI agents.
The modern developer is transitioning from an individual contributor to a manager of digital employees.

We believe that future 10x engineers will be managing, verifying, and reviewing 10+ tasks simultaneously done by AI.

Austin Wang, Co-creator, Manaflow AI · Austin Wang on LinkedIn
Key Takeaways

The Notification Soup Problem

The terminal multiplexer was invented for human engineers to manage multiple servers. Now humans need to manage multiple AI engineers. Running five instances of Claude Code in traditional tmux panes creates an unreadable soup of asynchronous notifications and lost context. When agents run concurrently, they all eventually pause to wait for human input. In a standard terminal setup, these prompts get buried. Developers are forced to constantly cycle through panes just to find blocked agents.

A tangled ball of ticker-tape contrasted with neatly organized parallel glass tubes containing straight ticker-tape. This shows the difference between traditional terminal chaos and cmux organization.
Traditional multiplexing tangles asynchronous agent outputs, while cmux isolates them into manageable parallel streams.

A Dashboard for Digital Employees

To solve this, the Manaflow AI team built cmux. It replaces horizontal tabs with a vertical sidebar that acts as an agent metadata dashboard. This sidebar shows the current git branch, PR status, and the latest notification text for each active agent. When an agent blocks and needs human input, its pane gets a glowing blue ring. The developer can see exactly who needs help without hunting for it.

Hedcut portrait of Austin Wang

Native Speed with libghostty

Building a terminal that can handle five parallel LLMs spitting out tokens requires serious performance. Instead of building another sluggish Electron wrapper, cmux is a native AppKit macOS application. It embeds libghostty for GPU-accelerated rendering. This allows it to ingest massive walls of token output without dropping frames or spiking CPU usage.

The asynchronous agent loop. cmux uses visual top-level indicators to proactively alert the developer when a specific agent sandbox requires human intervention.

Giving Agents a Browser

cmux does not just host the terminal prompt. It spawns agents in isolated Docker containers or cloud sandboxes. Crucially, it includes a scriptable split-pane browser. Agents need to see what they build, and humans need to verify it. By integrating the browser directly into the multiplexer, developers can instantly see the UI changes an agent just coded.

A magnifying glass held over a blueprint, where the lines inside the lens become a 3D building. This represents the integrated browser turning code into verifiable UI.
The integrated split-pane browser allows developers to instantly verify the visual output of agent-generated code.

cmux helps software engineers context switch fast between tasks, safely deploy AI agents in cloud sandboxes or local Docker containers, and verify AI-written code through preview environments.

Austin Wang, Co-creator, Manaflow AI · Austin Wang on LinkedIn

Multiplexers Compared

FeatureLegacy Multiplexers (tmux)Agent Multiplexers (cmux)
Primary UserHuman EngineerAI Agent
State IndicationBuried text outputVisual UI badges and rings
EnvironmentShared local hostIsolated sandboxes
NavigationHorizontal tabs / panesVertical metadata sidebar