What Happened
A developer launched ChatML, an open-source desktop application for managing multiple Claude Code sessions simultaneously. The pitch: instead of running one Claude Code agent at a time, run several in parallel, each working in its own isolated git worktree on separate features or fixes.
The developer behind it claims they didn't write a single line of code for 45 days while building the app. They described what to build, ran multiple Claude agents in parallel with isolated worktrees, and spent their time reviewing diffs and making architectural decisions. The result is the app itself - built by the workflow it's designed to enable.
ChatML runs on macOS 13+, Windows 10+, and Linux. Key features include parallel session management with a unified dashboard, sub-agent spawning where agents can delegate tasks to child agents, built-in code review at multiple depth levels (quick, deep, security), real-time cost and token tracking with budget limits, MCP server support for extending capabilities, and Linear integration for issue tracking. It's built on Anthropic's Claude Agent SDK.
The app is free during beta and fully open source.
Why It Matters
Running one Claude Code agent is productive. Running three or four on different parts of a codebase simultaneously is a different workflow entirely. The bottleneck shifts from "how fast can the agent write code" to "how fast can I review and merge the output."
Git worktree isolation is the right architectural choice here. Each agent gets its own working directory and branch, so there are no merge conflicts during parallel development. You review each agent's work independently and merge when ready. This is the same pattern that large engineering teams use with feature branches, but with AI agents instead of humans.
The cost tracking feature matters more than it might seem. Running multiple Claude Code sessions burns through API credits fast. Being able to set budget limits per session and monitor token usage in real time prevents surprise bills - a real concern when agents are working autonomously.
For solo developers and small teams, this kind of tool multiplies output without multiplying headcount. You're effectively running a small engineering team where you're the tech lead reviewing PRs from AI contributors.
Our Take
The "I didn't write code for 45 days" claim is the kind of thing that gets attention on Reddit but deserves scrutiny. Directing multiple AI agents, reviewing their output, making architectural decisions, and handling the inevitable cases where agents produce wrong or conflicting code is still work. It's different work, but it's not "the AI did everything."
That said, the tool itself addresses a genuine gap. Claude Code is powerful but single-threaded by default. The official way to run parallel sessions requires manual terminal management and git worktree setup. ChatML wraps that in a proper UI with the monitoring and controls that make multi-agent workflows practical rather than experimental.
The competitive landscape here is filling up fast. Composio's Agent Orchestrator, OpenWork, and Anthropic's own multi-agent documentation all point to parallel agent orchestration becoming a standard pattern. ChatML's advantage is that it's free, open source, and shipping now across all three desktop platforms.
If you're already using Claude Code daily, this is worth trying. The parallel workflow takes some adjustment - you need to think about task decomposition differently when you can run four agents at once - but the productivity ceiling is meaningfully higher than single-agent development.