- Published on
Claude Code's new Mods system lets developers rewrite the AI coding tool from the inside
- Authors
- Name
- DeepHorus
Anthropic has released Mods, a plugin system that lets developers modify Claude Code from the inside. JavaScript or TypeScript functions can hook into tool calls, user prompts and interface rendering. Developers may add panels beside the chat, intercept tool calls and introduce new commands.
Some built-in features already use the architecture, including the /diff command. Anthropic says this shows Mods can function as core middleware rather than external integrations. The company also publishes sample Mods on GitHub.
Mods run in Claude Code’s command-line interface and desktop app, with partial support in its VS Code extension. They operate with the user’s permissions and are not sandboxed. Anthropic warns users to install them only from trusted sources, while organizations can control which ones may load.
The first official Mod plugin, “You Should Know,” starts a separate agent to watch Claude’s output. It sends a “Heads up” message when it spots important information users might have missed. Users can activate it with /plugin enable cc-plugin-you-should-know@builtin.
The system could help developers tailor Claude Code to specific workflows and interface preferences. Event-level access gives Mods broad influence over prompts, tool calls and displayed content. That power also increases security responsibility because plugins inherit the permissions of the person running them.
Anthropic’s warnings and organizational controls acknowledge that unsandboxed extensions require careful management. Trusted-source guidance can reduce exposure to harmful code, and allowlists can restrict what teams deploy. Yet the access that enables customization also makes plugin review essential.
Current support spans the CLI, desktop app and part of the VS Code extension. Available patterns range from custom side panels and commands to intercepted tool calls and an independent review agent. Future adoption will depend on how users balance flexibility against the risks of running code without a sandbox.
