AI - Claude Code gets custom themes, a more proactive Projects coordinator and an experimental Code Mode
Anthropic did not exactly slow down over the weekend 🤣
After a week of Claude Meetings and self-hosted Managed Agents, Claude Code got three updates of its own: custom themes for the desktop app (yes, including animated backgrounds and snow 😊), a more proactive coordinator in Projects, and an experimental Code Mode that lets Claude write a bit of JavaScript to run many tool calls in one go.
One of them is pure fun, one is about managing agents, and one could make your usage limits last longer. Let me go through them in that order 😉
Custom themes for Claude Code on desktop
Custom themes are built on mods, the plugins that change how Claude Code looks and behaves from the inside. A mod can now change the colours of the desktop app and draw an animated backdrop behind the conversation, and because it is just a mod, the backdrop can react to what is happening in your session.
Official announcement
You can now build custom themes for Claude Code on desktop with mods! Change the colors, add an animated background, and make it your own.
Just tell Claude how you want it to look, or get started quickly with the theme-examples plugin. We’d love to see what you make!
What is new?
According to the Claude Code documentation and the plugin’s README:
- Colours and backdrops from a mod: a mod calls
$.theme.setto set the app’s colours for light and dark mode and an animated backdrop behind the conversation.$.theme.clear()takes it off again, and so does/plain. - The theme-examples plugin: Anthropic’s official plugin has twenty animated scenes, each made for a light or dark window, from a lagoon seen from above and a koi pond to a departure board and a starfield in braille dots.
/custom-themeshows a card with all of them and a preview button, and you can apply one to a whole project or to a single session. - A bit of fun on the side: the same plugin adds
/fireworks, and a show also goes off when Claude says “merged” (you can turn that off with/fireworks auto off) 🤣 - Or just ask Claude: Claude Code has a built-in
plugin-authoringskill, so you can describe the theme you want and let Claude write the mod for you. - Accessibility: the scenes stop moving when your system asks for reduced motion.
Themes need the Claude desktop app (the docs say the mobile app can also draw backdrops). Mods themselves also work in the CLI and the VSCode extension, but the terminal does not draw a backdrop. $.theme.set needs Claude Code v2.1.319 or later, and mods are not available in WSL sessions in the desktop app.
To try it, install the plugin with claude plugin install theme-examples@claude-plugins-official (or /plugin install theme-examples@claude-plugins-official in a session), then run /custom-theme in the Code tab of the desktop app.
A mod is code that runs with your permissions, so only install mods from authors and marketplaces you trust. Admins can control which mods load through managed settings.
From the people who built it
Custom themes are just mods so your background can even react to your actual session!
Here I just asked Claude to make it snow as my context fills up (the mod hooks into turn.step) but you can get really creative with it ☃️
Snow that piles up as your context fills is honestly a better context indicator than a percentage 🤣
Showing off more Claude Code custom themes
What people are saying
Translated from Korean:
🎨 You can now freely customise themes in Claude Code desktop!
Install the official Anthropic plugin = theme-examples, let us go
In the terminal, it is like this.
> claude plugin install theme-examples@claude-plugins-official
Not only can you change the colours, you can now even put an animated background behind the chat window.
You can use the 20 different samples, or you can tell Claude “make it with this kind of feel”.
I installed it too, and it completely changes the mood, it is fun lol
Anthropic has been really focused on the desktop app lately and has been pushing out a huge number of updates.
So that we use this tool more often, and for longer..
I don’t think most people realize how sick Claude Code is now
You can literally prompt it to change ANY part of the interface you want
I asked it to build a football theme since it’s Sunday. I now have live scores literally built into the UI of Claude.
Now I can lock in all day and miss nothing going on in the NFL
Other awesome ideas you can implement:
> A theme that shows you your usage at all times
> A theme that actively changes based on how many tokens you’ve burnt that day so you can feel like you’re building streaks
> A theme that shows you analytics on the live project you’re building
> A theme that shows how many visitors you currently have on your site, keeping you motivated as you build
> A theme that shows the status of all your other projects so you can keep tabs in one placeSo many possibilities. You NEED to be playing around with this (give this post to your agent to take some of these ideas)
Alex’s ideas show where this gets interesting. A theme is no longer just a colour scheme; it is a little dashboard behind your conversation. Just remember that live scores mean a mod making network requests with your permissions, so the “only install what you trust” rule very much applies to anything you did not write yourself 😉
Projects in Claude Code: a more proactive coordinator
Projects let you hand Claude a stream of related tasks in one conversation. Claude, as the coordinator, starts a thread for each task (usually a cloud session), and the threads run in parallel while you check on them from the project. This update makes the coordinator much more proactive.
It is a similar idea to Cursor’s Projects coordinator agent, which I wrote about earlier.
Official announcement
We’ve made some big updates to Projects in Claude Code!
In a project, Claude coordinates your tasks across parallel threads. The coordinator is now much more proactive. It answers your questions in the project chat and gives you updates on the status of your ongoing threads.
What is new?
- A more proactive coordinator: according to Anthropic, it now answers your questions in the project chat and gives you updates on the status of your threads without you having to go and look.
- Separate models for the coordinator and threads: the docs list Coordinator model and effort and Thread model and effort in Project settings > General. A new project runs Opus everywhere, with high effort for threads and low effort for the coordinator, so this is a handy way to save usage, for example with a smaller model for simple threads.
- You can tell it how to coordinate: for example, “Run at most two threads at a time”, “Only post when something finishes or is blocked” or “Give me a status update on every thread”. Claude saves these to project memory, although the docs are clear that they are instructions, not enforced settings.
Projects are in public beta on Pro, Max and Ultra plans and are rolling out gradually; they are not offered on Team or Enterprise plans yet. They are available at claude.ai/code, in the desktop app and in the Claude mobile app, but not in the terminal CLI, the VSCode extension or the JetBrains plugin. If you cannot see Projects in the sidebar (check under More too), the rollout has not reached you yet, and you can join the waitlist.
Cloud threads need your code on github.com with the Claude GitHub App installed. A new project runs Opus at high effort for every thread, which uses your plan fastest, so check Thread model and Thread effort in Project settings > General before sending it a big batch of work.
What people are saying
TestingCatalog spotted the model settings and a few other changes:
ICYMI 👀: You can now configure different models for Coordinator and Threads in Claude Code Project settings.
Projects in overal got many small upgrades recently, including a possibility to create custom views and assign atrifacts to them.
This is Conway 🤖
The separate coordinator and thread models are confirmed in the docs. The custom views and “Conway” name are TestingCatalog’s own findings, and I could not find them in the official documentation.
Anthropic is coming after t3code 😅
Not everyone is happy yet
Chris Laupama has been using it and has some very fair criticism:
Hey @AnthropicAI
- The orchestrator needs to keep an eye on the threads more and bring more to my attention, I’m finding that I’m having to check the threads myself to find out what’s happening a lot.
- The threads go idle a lot. They may be in the middle of a list of tasks and just randomly stop even though it clearly has more to do, the orchestrator should be keeping an eye on this.
- Give me more control over permissions, having no way to allow threads to have more freedom is a huge time waster as it could ask me for permission but the orchestrator has no idea so I never get notified unless I constantly keep an eye on the threads myself.
Please fix these issues 🙏🏽
The permissions point matches the documentation. Threads run in auto mode where the model supports it, but when a thread does need approval, the prompt sits inside that thread, and telling the coordinator to go ahead does not reach it. Cloud threads also do not read the permission rules in a repository’s .claude/settings.json. The docs do say that the desktop app shows a notification when a thread needs your input, so it is worth checking that notifications are on for the project. Still, a coordinator that is supposed to coordinate should really know when one of its threads is stuck, so I hope this is next on the list 😉
Code Mode (experimental)
The third update is the most technical, but potentially the most useful. Normally, Claude asks for one tool call (Bash, Read, Edit and so on), reads the result, then decides what to do next. With Code Mode, Claude writes a short piece of JavaScript that makes several tool calls in one step: it can run commands in parallel, use one result in the next call and return only the part of the output Claude actually needs to read.
Official announcement
There is no official X post for this one; it arrived in the Claude Code v2.1.323 changelog:
Added Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools; see https://code.claude.com/docs/en/code-mode to turn it on, or to turn it off for an organization
Run tool calls from code with Code Mode
Code Mode changes how Claude uses its tools. Without Code Mode, Claude asks for tool calls such as Bash, Read, or Edit directly and reads each result before it decides what to do next. With Code Mode, Claude writes a piece of JavaScript that makes the tool calls, and Claude Code runs that code in your session.
What is new?
- Fewer round trips: the docs suggest it for a change that repeats across many files, a search whose results need filtering before Claude reads them, and several independent commands that can run at the same time.
- You still see what it does: the conversation shows the tool calls, such as an edit’s diff, rather than the code. Press
Ctrl+Oto open the transcript viewer and expand a Code row to read the JavaScript. - Your rules still apply: every tool call the code makes still goes through your
denyandaskpermission rules andPreToolUsehooks. In auto mode, the classifier reviews each piece of code rather than each tool call, so the docs recommend keeping rules or hooks for calls you must control. - Time limits: code that Claude waits on stops after ten minutes; in auto or
bypassPermissionsmode, code can run in the background for up to two hours. - Admins can block it: set
CLAUDE_CODE_EXPERIMENTAL_CODE_MODEto0in managed settings. You cannot switch it on for users that way, and a repository’s own settings cannot turn it on for you either.
Code Mode is experimental and off by default, and it may change or be removed in any release. It has quite a few requirements:
- Claude Code v2.1.323 or later, the native build, in an interactive terminal session. It does not work in the VSCode panel, the desktop app, cloud sessions,
-por the Agent SDK. - A Claude subscription or Console account, not a gateway, Amazon Bedrock, Google Cloud or Microsoft Foundry, and not in organisations with the HIPAA configuration or zero data retention.
- Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2. On Windows, Git for Windows must be installed.
- It will not start with an API key helper, agent teams, OpenTelemetry export or
DISABLE_TELEMETRY, and it does not yet work with an installed mod that checks tool calls. - You cannot use Remote Control, Claude in Chrome or computer use with it, and you can only resume a Code Mode session in the terminal with Code Mode on.
To try it on Windows, run this in PowerShell:
| |
On macOS, Linux or WSL, run CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude, or add the variable to the env block in ~/.claude/settings.json to turn it on for every session. Confirm with Yes, turn on Code Mode, and you should see Code Mode (experimental) is on near the top of the session.
What people are saying
Claude Code now has Code Mode too.
pi shipped this a while ago as codemode.
Instead of calling tools one by one, Claude writes a short JavaScript snippet that runs them for it. One script can search files, filter results, make edits, and run tests.
The intermediate results stay inside the script. Claude only reads the output it needs, so you save a lot of tokens on tool-heavy tasks.
It’s experimental. Update to v2.1.323+, then start it with:
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude
Kai’s summary is spot on, and it is a good reminder that this idea is not new. Other agents have been doing “code instead of tool calls” for a while; it is just nice to see it in Claude Code itself.
Translated from Chinese:
Holy cow, turn on this feature and Claude can save 37% of your tokens!
Upgrade Claude Code to v2.1.323,
Upgrade Claude Code to v2.1.323,
Upgrade Claude Code to v2.1.323,
then set CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude
Confirm that “Code Mode (experimental) is on”, and you are done!
The 37% figure is not something Anthropic states anywhere in the documentation, so treat it as one person’s claim. How much you save will depend heavily on how tool-heavy your work is.
Let Claude check whether Code Mode is worth it for you
Vox shared a really practical prompt that asks Claude to review your own sessions and settings and find where Code Mode would help, and what would stop it from starting:
Anthropic just published a guide to Code Mode in Claude Code. It tells you that when a job takes many tool calls in a row, Claude should 𝘄𝗿𝗶𝘁𝗲 𝗼𝗻𝗲 𝗽𝗶𝗲𝗰𝗲 𝗼𝗳 𝗰𝗼𝗱𝗲 𝗮𝗻𝗱 𝗿𝘂𝗻 𝘁𝗵𝗲𝗺 𝗮𝗹𝗹 𝗮𝘁 𝗼𝗻𝗰𝗲. You can ask Fable 5.2 to use it to find those jobs in your project.
Normally, Claude stops after every tool call to read the result. With Code Mode on, it finds the file, edits it and runs the tests in one go, then reads only the part of the output it needs. Fewer round trips, so 𝘆𝗼𝘂𝗿 𝟱-𝗵𝗼𝘂𝗿 𝗮𝗻𝗱 𝘄𝗲𝗲𝗸𝗹𝘆 𝗹𝗶𝗺𝗶𝘁𝘀 𝗹𝗮𝘀𝘁 𝗹𝗼𝗻𝗴𝗲𝗿.
Send this prompt and the link below to Fable 5.2 👇
“Read this guide, then review my user and project Claude Code settings, my environment variables, and the transcripts of this project’s last 20 sessions in ~/.claude/projects. Find three kinds of things: work I keep repeating that takes many related tool calls, like the same change across many files, searches whose results need filtering, or commands that could run at the same time; settings that stop Code Mode from starting, like an API key helper, agent teams, OpenTelemetry export or DISABLE_TELEMETRY; and, among my own hooks and those that come with plugins, PreToolUse hooks that would refuse tool calls in Code Mode and UserPromptSubmit hooks that would hold back code output.
For each one, point to the session, setting or hook, explain its impact, and suggest the smallest change needed. Rank the work you found by how many round trips Code Mode would save, biggest savings at the top. Show me the recommendations first. Don’t modify anything yet.”
I like that the prompt ends with “Don’t modify anything yet” 😊 Reading your recent transcripts and settings is a lot of context to hand over, so it is worth running on a project where you are comfortable with that, and it will of course use some of your limits to save some of your limits 🤣
My take
Three very different updates, and I like all three for different reasons.
Custom themes are pure fun, but there is more to them than looks. Because they are mods, a theme can show your context, your usage or your CI status right behind the conversation, and Lydia’s snow is a great example of that. Just treat mods like any other code you run: if you did not write it, check where it came from.
The Projects update is the one I am most curious about long-term. A coordinator that keeps an eye on parallel threads is exactly what you need once you have more than two or three agents running, and being able to put a cheaper model on simple threads is a good touch. But Chris’s criticism shows it is still a beta: if a thread can sit waiting for approval without the coordinator knowing, you are back to checking every thread yourself.
Code Mode is the one that could save the most usage, at least for tool-heavy work. It is a pity that, yet again, it does not work in the VSCode extension (or the desktop app), only in the terminal. I have not tried any of these yet, but Code Mode is first on my list, and it is good that Anthropic made it a deliberate opt-in with a managed setting to block it, rather than switching it on for everyone 😉
Useful links
- Claude Code Docs: Run tool calls from code with Code Mode
- Claude Code changelog
- Claude Code Docs: Let Claude coordinate ongoing work with Projects
- Claude Code Docs: Run agents in parallel
- Claude Code Docs: Mods overview
- Claude Code Docs: Draw in the interface with a mod
- Claude Code Docs: Manage mods for your organization
- theme-examples plugin on GitHub
#Blog #AI #Claude #Claude Code #Anthropic #Claude Desktop #Mods #Themes #Claude Code Projects #Code Mode #Agents