Claude Code Adds a /fork Command: Branch Your AI Coding Session Git-Style

Anthropic has added a new command, "/fork," and a refreshed command-line interface (CLI, the text-based screen used to operate the tool) to its AI coding tool, Claude Code. The addition was reported roughly ten days after Microsoft wrapped up its Build 2026 developer conference (held June 2–3 in San Francisco).^[1] It is not a flashy new model, and at first glance it looks minor, but because it maps directly onto a workflow engineers repeat every day, its impact on usability is larger than its size suggests.

What the "/fork" command does

"/fork" branches the session you are currently working in with Claude Code into a second, parallel copy on the spot. A session accumulates all of its context: the problem you are tackling, the files involved, and the instructions you have given so far. When you fork it, you get a duplicate that inherits that context, letting you try a different approach without breaking the original.

If the new direction works out, you keep that branch. If it does not produce the result you wanted, you discard it and continue from the untouched original session. Until now, trying an alternative meant re-explaining the premise each time or manually setting aside your work in progress. With forking, that switch takes a single command.

A familiar idea borrowed from Git

The concept is lifted almost directly from branching in Git, the version-control tool widely used in software development. In Git, you cut a branch to try out a new feature without affecting the main codebase. Any developer who uses branches day to day will grasp the idea of forking immediately.

By matching a working pattern people already trust, Anthropic keeps the burden of learning new commands low. Whether you want to compare two prompts (the instructions you give the AI) or test two different implementations, the barrier to experimentation drops sharply.

How it pairs with sub-agents

The "/fork" command makes the most sense read alongside the "sub-agents" feature added a little earlier. Sub-agents let a central session call up several helper AIs that split a single task and work on it concurrently.

Where sub-agents "divide one approach and process it in parallel," forking is responsible for "running multiple approaches in parallel." Cut a branch to explore an alternative, delegate the fine-grained work, and keep the best result at the end. Claude Code is moving away from the single linear chat that characterized the first generation of AI coding assistants and toward something closer to how engineering teams actually operate.

Why this is a developer-adoption story

These unglamorous improvements strike at the heart of a core question: how far can you trust an AI across long, multi-step work? The longer a task runs, the more the odds of the AI heading down a wrong path compound, and once a single linear session goes off the rails, rescuing it is hard work. Forking turns that risk into a cheap experiment, cut a branch, try it, and throw it away if it fails. Because the cost of failure is bounded, developers can hand the AI more ambitious work.

Behind this is the fact that Claude Code is the fastest-growing product in Anthropic's history. One analysis estimated that around 4 percent of all public GitHub commits (the change history of code) worldwide were, at one point, authored by Claude Code. Developers adopt it on their own, and corporate spending follows, a bottom-up spread that has lifted Anthropic's business standing. Features like forking and sub-agents are a move to further solidify that foothold.

It is worth noting that Claude Code's features have been updating on a near-weekly cadence, so the exact command syntax and behavior are best confirmed against Anthropic's official documentation.

Summary

The "/fork" command brings the "branching" that developers take for granted in Git into the AI conversation itself. Because you can try alternatives while preserving the original session, redoing failed attempts on long tasks becomes lighter, and the range of work you can hand to the AI widens. Combined with the sub-agents added earlier, you can run delegation and trial-and-error at the same time. It is far less conspicuous than a new model announcement, but it is a practical step that brings AI coding closer to being a tool you can confidently delegate to.