Claude Code now supports mods: TypeScript extensions that can change behavior, customize the UI, or swap in new features. Mods ship inside plugins, installed via /plugin in the CLI or desktop app, and can be written manually or generated by Claude itself.
Features already shipped in Claude Code, including the /diff command and AGENTS.md support, were built internally using the mods system. Sample mods are being added to the Claude Code Playground repo as starting points.
Mods support spawning forked agents mid-session, which opens up patterns like running a custom classifier on every turn to decide whether to write a memory. A plan-mode mod is also in progress.
Framing the mods launch as part of a broader shift toward malleable software, the post argues extensibility should be a first-class citizen in any tool, not an afterthought.
A detailed thread from an early Claude Projects user covering how the main chat, threads, effort levels, local compute, and async notifications all work together. Early access available via DM.
The main chat in Claude Projects runs at low effort by default and delegates work to threads, which default to Opus 5.5 at medium effort. Each thread is a full Claude instance, not a transient helper, and persists for direct conversation.
Claude Projects is in public beta for Pro and Max subscribers, rolling out gradually. The launch post covers the thinking behind the feature.
Some tasks require local compute; users can connect a folder in the Claude desktop app, and that thread runs locally while other threads continue in the cloud.
Threads in Claude Projects run in the cloud by default, so work continues after closing your laptop. The Overview shows which threads finished and which are waiting, and users can check in from their phone.
Anthropic.com/engineering is live as a dedicated home for developers, featuring engineering deep dives, Claude Code and API guides, tips from the teams building Claude, and some easter eggs.
Any pane in the desktop app, including the diff view and terminal, can be detached into its own window and dragged to a second monitor while Claude keeps working. Docking it back is one drag away.
A user reported that cloud agent auto mode keeps interrupting for approval on small edits even with 'dangerously bypass permissions' enabled. An Anthropic engineer acknowledged it and is investigating as a likely bug.
In response to a user confused about configuring auto mode for cloud sessions, an Anthropic team member clarified that cloud sessions can be selected from the mode dropdown like any other session, and asked whether the user wanted to set custom auto mode rules instead.
The desktop app bundles its own copy of Claude Code and updates it independently from the terminal version. Right now it's a few versions behind, running 2.1.284 while the terminal has already pulled 2.1.287.
The lead on Claude Code calls the new mods system a fundamental shift: instead of everyone using an identical tool, you can now prompt Claude to build a mod that changes how it works and looks, then share that mod as a plugin for others to install.
Token Weather is an example mod that displays a live forecast of your context window fill level alongside a sparkline of usage across your last 12 turns.
The Blast Radius example mod intercepts dangerous shell commands like rm -rf, git reset --hard, and force pushes before execution, showing what would be affected in a side pane.
A new built-in mod called Replay Theater records every file change Claude makes during a turn. Run /replay to step through the diffs one at a time in a docked side pane.
Mods run with the same system permissions as Claude Code itself, so the team is advising users to install only from trusted sources. Private mods stay local; useful ones can be submitted to the Claude directory.
