Anthropic shipped something two weeks ago that we haven't stopped thinking about.
It's called Dispatch. It's part of Cowork, the Claude agent that runs on your desktop and works directly in your files. The short version: you can send tasks to your computer's AI from your phone, and it handles them while you're away. But the why this matters part takes some explaining.
Nobody outside the AI-tools world has noticed yet. Give it a few weeks and the LinkedIn crowd will catch up. Consider this an early look.
What Cowork Actually Is
Cowork launched in January as part of Claude Desktop, the Mac app that puts Anthropic's AI on your local machine. The difference from every chatbot you've tried: it doesn't just generate text. It does things inside a folder you designate.
You point Claude at a folder. You describe what you want. It reads your files, creates new ones, rewrites existing ones, builds spreadsheets from scattered documents. It does this on its own, only asking you for input if it gets stuck. You describe the task in plain English, the way you'd explain it to a new colleague.
Anthropic's own examples give you the idea: turn a pile of receipt screenshots into an expense spreadsheet, turn a folder of meeting notes into a draft board report, reorganize a messy shared folder so people can find things.
Then They Added Dispatch
On March 17th, Anthropic released Dispatch.
Dispatch creates a persistent connection between Claude Desktop on your computer and Claude on your phone. So you leave the office, remember you need a report summarized, open your phone, send the request, and your laptop handles it. The output is there when you wake up.
They shipped task scheduling at the same time: recurring tasks, on-demand queues, the ability to stack jobs and let Claude work through them in parallel.
We've been waiting a while for something specific: an AI agent that non-technical staff could use, unsupervised, on real work, with real files. This is the first thing that clears that bar.
What This Looks Like in Practice
"AI agents" has been vaporware for most people, so let's talk about actual work.
A grants manager at a 20-person housing nonprofit. Every month she gets reports from eight program sites: Word docs, PDFs, a few spreadsheets. She needs a one-page summary of each for the board package. It takes most of a Friday.
With Cowork, she drops the files in a folder, writes one instruction ("summarize each document into a one-pager for a non-specialist board audience, pull out numbers and outcomes"), queues it Thursday evening, and checks her phone in the morning.
Or a municipal policy team writing a council report. Three analysts with overlapping notes in a shared folder, nobody sure who wrote what. Point Cowork at the folder: "draft a brief that pulls these notes together by theme, flag contradictions." By morning there's a first draft that would have taken a full day to assemble by hand.
We're setting up both of these workflows for clients right now.
Why This One Is Different
Two things separate Cowork from previous attempts at AI file agents.
First, sandboxing. Claude can only access the folder you explicitly mount. It runs in an isolated environment on your machine. Nothing else is exposed. For public sector teams with data handling concerns, this matters. It's not ambient AI watching your whole computer. It's a bounded tool with one designated workspace.
Second, there's no technical setup. You install Claude Desktop, subscribe to a plan, and it works. No API keys, no pipelines, no developer needed. We watched a 58-year-old program director who had only used ChatGPT a few times get Cowork running on her own in under ten minutes.
The bottleneck for AI adoption at most small orgs has never been "the AI isn't good enough." It's been "we don't have anyone technical to set it up." Cowork sidesteps that problem entirely.
The Limitations
We like this tool a lot but we're not going to pretend it's perfect.
It's Mac-only right now. Anthropic has mentioned Windows support but hasn't shipped it. Most government offices and a lot of nonprofits run Windows, which means you'll need to wait or find a workaround.
Pricing is Claude Max, which runs $100 to $200 per month per seat. Not cheap, especially for lean organizations. But if a grants writer saves eight hours a month, the subscription pays for itself in the first week.
It's also still a research preview. Tasks sometimes stall. Complex multi-document work needs clear, specific instructions or the output won't be useful. Start with one simple workflow, tweak the prompt until it's reliable, then expand from there.
Where to Start
Pick the workflow that eats the most time and involves files already on someone's computer. That's your test case.
For nonprofits, grant report compilation is the most common starting point we see. Policy teams tend to start with turning meeting notes into briefs. Ops-heavy orgs usually go for expense processing. The pattern is the same: documents in, document out.
Set up a folder, write the instruction, test it on one real task before you rely on it. You'll know within 30 minutes whether it fits.
We've been working with Dispatch for two weeks. If you want to see it running before you commit to anything, or if you want someone to set up the folder structure and prompts for your team, that's what our sprints are for. We get the tool working in your actual workflow by the end of the week. Get in touch.