Claude Dispatch: My Phone is Now my Remote
I was on a walk today, my desktop sitting at home with Claude Desktop running. I opened Claude on my iPhone and typed: "Read the plans in my docs folder and analyze what’s left to be done before implementation can start."
A couple of minutes later, Claude replied with a full critique, next steps, and P1s to file. My computer had done all the heavy lifting, and I could continue the conversation as long as I needed to from my phone.
I honestly couldn’t believe how seamless it felt. It felt very OpenClaw-esque, but now from Anthropic. It's called Claude Dispatch
What Dispatch Does
Dispatch is a feature inside Anthropic's desktop agent product, Cowork. It lets you send messages from the Claude mobile app to the Claude agent running on your desktop. The conversation continues across devices, so you can start a task on your phone and pick it up later from your computer in the same thread.
Your phone becomes a remote control for an AI agent sitting at your desk.
The agent can use whatever Cowork already has access to: local files, browser sessions, connected services, and MCP integrations. Dispatch doesn't add new intelligence; it adds a new way to reach the intelligence already running on your machine.
Setup is simple: update Claude Desktop and the mobile app, enable Dispatch in Cowork, and grant the access you want.
Local Execution
The important detail is where the work happens. When you send a Dispatch message, Anthropic isn't spinning up a cloud VM. The work runs on your own computer, which means Claude can operate inside your real environment: filesystem, installed apps, live browser sessions, dev setup, and custom MCP servers.
Most of the industry is moving in the other direction. Cloud-based agents are easier to operate; your laptop doesn't need to stay awake, there are fewer local dependencies, and the environment is cleaner. But cloud agents have a ceiling. They can browse the web in a sandbox, but they can't reach into the messy, valuable, real context of your machine.
Anthropic is betting the mess is where the important things live.
Cloud execution works better for tasks that mostly live on the web: research, browsing, and public information gathering. Local execution works better for tasks that depend on your real environment: local files, browser sessions, installed tools, dev setups, and custom integrations.
The Rough Edges
Dispatch is still early, so you might notice some rough edges.
If your desktop goes to sleep, the task stops. If Claude Desktop isn't open, nothing runs. Notifications are basic. You only get one thread. The main workflow is solid, but everything around it still feels unfinished.
The bigger concern is security. Dispatch creates a chain: your phone sends a message, Anthropic routes it to your desktop, Claude interprets it, then acts on your local files, browser, and connected services. A lot of trust flows through that pipeline.
Anthropic's mitigations include permission checks to prevent destructive actions, content scanning to detect prompt injection, sandboxed code execution, and training to help Claude refuse bad instructions. Those are reasonable, but the core problem is the same one every agent platform faces: prompt injection isn't solved. If an agent reads malicious instructions hidden inside a webpage or document, it may treat them as legitimate. Local access raises the stakes because a cloud agent makes mistakes in a sandbox; a local agent makes mistakes in your actual workspace.
The golden rule: only give the agent access you would be okay with if something goes wrong.
Same Pattern for Developers
Dispatch isn't just a one-off. It's part of a bigger push to keep agents connected to the environment where they actually work. Anthropic released the same idea for developers with Claude Code Remote Control. Start a session in your terminal, pair it with your phone, and keep controlling it remotely. The terminal, project files, and toolchain all stay local. Your phone is the handle. Under the hood, it uses the same relay architecture as Dispatch: your machine makes outbound HTTPS connections to Anthropic's API, so it works behind any firewall without port forwarding. We wrote a detailed walkthrough of how Remote Control works, including the security model, session gotchas, and how it compares to Claude Code on the Web.
Anthropic Wasn't First
The idea didn't start at Anthropic. OpenClaw had already shown the shape of it: run an agent locally, talk to it through a messaging app, let it work on your machine, and get results back wherever you are. OpenClaw pushed the concept further in some directions too, with broader messaging support, model flexibility, scheduled tasks, more integrations, and proactive behavior.
The open-source world had proved that "text your agent from anywhere" was a real use case before Anthropic productized it. Anthropic's contribution is to validate the pattern and tie it tightly to Claude's ecosystem.
What This Signals
Anthropic's bet is that the most useful agent isn't just a better chat window or a cloud browser. It's a worker that lives in your real environment, can be reached from anywhere, and uses your files and tools.
There are tradeoffs: more setup, more things that can break, and more trust needed. But this might be where the most useful agents end up. Cloud agents are simpler. Local agents are closer to your real work.
Dispatch is Anthropic's bet that being close to your real environment matters more to some users than pure convenience.
Links:
Comments
Sign in to join the discussion.
No comments yet
Be the first to share your thoughts!