Anthropic has confirmed the arrival of "Claude Mods" — custom mods for Claude Code built on TypeScript functions that run directly inside the agent process and can be attached to almost any of its events. Early access is already open via the CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 flag, and the community has released its first mods.

image
image
image

What Happened

On September 3, 2026, Claude Code developer @poteat opened issue #91870 in the anthropics/claude-code repository with a proposal for "function hooks" — TypeScript functions that run inside the Claude Code process rather than as external shell commands. On September 9, in a community update on the issue, Anthropic confirmed that the feature would ship "within a week," announced the name "Claude Mods" (a mod is a plugin based on function hooks), and stated that it can already be tested with the CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 flag in builds 2.1.269/2.1.272+. Anthropic published the source code for three built-in mods — sec-default, diff, and telemetry — in the mods/ folder of the main repository. The community has already released its first mods: sezaakgun/cc-arcade with Tetris, Snake, 2048, Doom, Minesweeper, and other games above the prompt, Arunjay4213/claude-mods with a context and cost tracker, and cc-storytime.

Context

From an architectural standpoint, this is not a model change or a new ML capability, but a classic event-based middleware pattern in the spirit of Express/Koa applied to an agentic CLI. Technically significant is the move of hooks inside the Claude Code process: mods do not run as external shell scripts but execute within the agent process and can render interfaces directly above the prompt in the terminal, reacting to almost any engine event, including UI rendering and input. The issue #91870 itself is titled "Mods - make Claude 10x more extensible."

Why This Matters for the Industry

For builders, this is a shift in the class of extensions: third-party TypeScript code runs inside the process and gains access not only to agent actions but also to UI rendering, input, and engine events. The role of administrative control grows — administrators will be able to programmatically restrict which system calls are allowed for installed plugins. For startups, this is an early and cheap distribution window in a new extension layer, although the business model has not been announced: there is no mod marketplace. The significance will be judged not by the idea itself — it is trivial as an engineering pattern — but by the stability of the Claude Code API and the size of the ecosystem that forms.

Why This Matters for Users

You can try it today: add the environment variable CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 to ~/.claude/settings.json in builds 2.1.269/2.1.272+ and install ready-made mods from GitHub — for example, games for pauses while Claude works, or context and cost trackers. The barrier to entry is low: the source code for the built-in mods sec-default, diff, and telemetry is available for study, and a "context and cost tracker" utility can be built in a few days.

What Is Still Unknown / Limitations

The feature is in early access: the API may change between Claude Code releases, so mods may break after updates; there is no SLA, and the materials do not provide data on latency, hook costs, or isolation of third-party code within the agent process. The phrases "full-fledged plugin ecosystem" and "a level of extensibility no major agentic CLI has had" are not supported: there are no comparisons with other agentic CLIs or benchmarks in the sources. The "10x more extensible" claim from the issue title is a marketing number without methodology, metrics, or a baseline.

Sources

Author

Look at AI, editorial team