Claude Mods Launches: Plugin Architecture Makes Claude Code Programmable

Anthropic has begun rolling out Claude Mods , a new plugin architecture for Claude Code that exposes the agent's internal execution paths to developers. The announcement came from Claude Code lead Boris Cherny on X, and the community has already produced a working Tetris game that runs directly inside the Claude Code interface — demonstrating that mods can inject custom React components and handle user interactions within the terminal UI.

From Function Hooks to Claude Mods

The feature originated as a GitHub proposal titled Function Hooks (issue #91870) on September 3. Unlike the earlier Hooks mechanism — which ran shell scripts at a few fixed lifecycle points (pre-tool-call, post-tool-call, session-start) and suffered from Windows compatibility issues — Function Hooks are typed TypeScript functions that sit directly in the execution pipeline. Multiple hooks chain together in registration order, forming a middleware stack analogous to Express or Koa. Each hook receives an event, can inspect or mutate it, decide whether to continue, and pass control to the next hook.

This design lets plugins:

Intercept and modify any tool call before execution

Wrap or replace React UI components, altering props or the rendered output

Capture low-level interactions such as button clicks

Enforce admin-defined capability boundaries for each plugin

On September 9, Anthropic branded the user-facing layer Claude Mods (a Mod = a plugin that uses Function Hooks) and opened the first built-in Mods — diff, sec-default, and telemetry — along with a test framework. Notably, the existing diff panel is being reimplemented as a Mod, and Anthropic states more core features will follow this pattern.

Community Reaction and Stress Testing

Developers on the GitHub issue have treated the early access as a live audit. Three concrete failure modes surfaced:

Silent hook timeout: A deliberately stalled hook failed silently, allowing the risky operation to proceed unchecked.

Tool substitution bypass: Intercepting a file write caused the model to switch to a different write tool, completing the same action because the interception provided no explanatory feedback to the model.

False-positive security checks: A security-review hook reported "all clear" for code that contained actual vulnerabilities.

These testers are production users who intend to integrate Mods into real systems, so they are effectively performing free pre-release security validation.

Industry Implications: Competition Moves to the Harness

The article argues that coding-agent differentiation is shifting from raw model capability to the harness — the surrounding infrastructure of context management, tool design, permission control, memory, sub-agent coordination, failure recovery, and organizational policy enforcement. Claude Mods hands developers programmable control over part of that harness.

If the ecosystem matures, selection criteria for coding agents will expand to include plugin ecosystem richness, granularity of control, enterprise policy maturity, and workflow embeddability. Anthropic is executing a classic platform play: open extension interfaces, validate them by rebuilding first-party features as Mods, then let the community populate the ecosystem. The long-term vision is that developers simply describe desired behaviors in natural language, and Claude writes the corresponding Mods itself.