Some weeks a tool ships one feature. Last week Claude Code shipped four, across five releases, and each one touches a different part of how we work. Remote Control lets you approve an agent's next step from your phone. Auto-memory means the agent starts remembering what it learned about your codebase without being told. HTTP hooks let other systems gate what it does. And a voice mode has started rolling out to a small slice of users.
Here is what each one is, how we are using it on client builds, and where we are holding back.
Remote Control: Approve From the Car Park
The problem it solves is familiar to anyone who runs long agent tasks. You start a migration or a large refactor, it will take forty minutes, and somewhere in the middle it needs a permission answered. You are either sitting at the desk waiting, or you come back to find it stalled at minute six.
Remote Control, released on 24 February as a research preview, connects the Claude app on your phone, or claude.ai in any browser, to a Claude Code session running on your own machine. The important word is "own". The code, the filesystem and the execution never leave your laptop. The phone is a window into a local session. From it you can read the conversation, send the next instruction, and answer permission prompts. If the laptop sleeps or the network drops, messages queue and the session reconnects when it comes back.
Starting it is one command, claude remote-control, or /rc inside an existing session, and a QR code to pair your phone.
For a small crew this is the feature of the month. Takeoff & Climb has a lot of forty-minute tasks. Being able to kick one off, walk to a client meeting and approve the two prompts it raises on the way is the difference between running three tasks a day and running eight.
Two limits worth knowing. It needs a Claude subscription login rather than an API key, so it does not work for sessions routed through Amazon Bedrock. And it is a preview on the Max plan first, so check your plan before you plan around it.
Auto-Memory: The Agent Keeps Its Own Notes
Since January we have shipped a CLAUDE.md with every client repository: the rules, the conventions, the things this codebase must never do. Keeping it current is a human job. Every correction we make in a session has to be written back by hand.
From version 2.1.59 Claude Code does part of that itself. It saves useful context to an automatic memory as it works, and you manage what it kept with the /memory command. A few days later that memory became shared across git worktrees of the same repository, which matters because we run one worktree per feature and did not want four agents learning the same lesson four times.
We like it with one rule attached. Memory the agent writes for itself is reviewed the same way code is. Once a week we read what it saved, promote the durable lessons into CLAUDE.md where the client's own developers will find them, and delete the rest. A memory nobody reads is how an agent quietly starts doing something you never asked for.
HTTP Hooks: Let Your Systems Gate the Agent
Hooks have been in Claude Code for a while. They run a shell command at points in the agent's loop: before a tool call, after a file edit, when a session stops. We use them to run the formatter after every edit and to block writes outside the project.
Version 2.1.63 adds HTTP hooks. Instead of a shell command, the hook POSTs JSON to a URL and acts on the JSON that comes back. That turns any service into a gate. A few uses we have set up on a client build already:
- A hook that posts each proposed pull request to a small service that checks it against the sprint's token budget and refuses the merge if the build is over.
- A hook that notifies the client's Slack channel when a feature branch passes its verification.
- A hook that asks our CI for the last green commit before the agent starts a risky refactor.
A security note from earlier in the week matters here: environment variables are only interpolated into hook requests if you list them explicitly, and hooks go through the sandbox's network allowlist when sandboxing is on. Good defaults. Keep them.
The same release added two bundled commands, /simplify for the post-feature cleanup pass we used to run as a custom subagent, and /batch for applying one change across many files in parallel.
Voice: Early, and We Are Waiting
Anthropic has also started rolling out a /voice command, live for about five percent of users with a wider rollout to come. You speak the instruction, the agent executes it. TechCrunch could not get answers on limits or on which provider handles the speech, and neither can we, because we are not in the five percent.
We will try it when it arrives. For dictating a spec while looking at a whiteboard it could be useful. For steering a refactor it is hard to imagine beating the keyboard. Ask us again in a month.
A Note on Concentration Risk
It was a loud week for Anthropic beyond the changelog. A dispute with the US Department of Defense made headlines, Claude topped the US App Store, and claude.ai and Claude Code had an outage while the API stayed up. None of that changes the features above. It does reinforce something we put in every architecture overview: the interface to a model provider should be a seam you can swap behind, and the consumer app should never be on a client's critical path. The API was fine. The people depending on the app were not.
What to Take From This
If you are a founder watching your build, the practical change is responsiveness. Long agent tasks no longer wait for someone to be at a desk, so the fortnightly demo gets more features in it. The memory and hook changes are about control: the agent learns, and other systems can say no to it. Both are what you should want from a crew using these tools on your product.
Building With an Agent That Remembers Your Rules
At Dev Force One, Claude Code is one of the two agents flying every Takeoff & Climb, and this week it got easier to supervise and harder to derail.
What that means for your build:
- More done per day: long tasks approved from a phone, not from a desk
- Rules that compound: a reviewed memory and a CLAUDE.md that your own developers inherit
- Systems that can say no: HTTP hooks tying the agent to your budget, your CI and your Slack
Want a build with that level of control? Book a Pre-flight call and we will walk you through it.