The Device in Your Pocket Is Your Most Underused Dev Tool
We built Seawork because the gap between where AI agents run and where you are shouldn't exist.
Here's a scenario that happened to us constantly before we built Seawork.
You start a Claude Code session to refactor a module. It's going to take a while — deep file traversal, multiple tool calls, maybe 20 minutes. You walk away to make coffee, take a meeting, or just move to another room. You come back 25 minutes later to find the agent froze waiting for permission approval at minute 3, and has been doing nothing ever since.
Or worse: it didn't freeze. It kept going, confidently, down the wrong path. Twenty minutes of compute and API spend building something you're going to delete.
This kept happening. And the "solution" was always the same: don't walk away. Sit there. Watch the terminal.
That's not a solution. That's babysitting.
The gap we're closing
Every serious developer has a powerful computer running their actual work. They also have a powerful computer in their pocket. These two things almost never talk to each other in a useful way.
We think that gap shouldn't exist — especially now that AI agents are doing real work that takes real time and needs real supervision.
Seawork is a mobile and desktop app that connects to your local development environment directly. Watch your agents in real time, approve permission requests the moment they come in, steer when something goes sideways. From your phone, from another room, from anywhere.
The design principle is simple: the agent's environment stays on your machine; the interface comes to you.
Why "your machine" matters
When we started building this, the easy path was to run everything in the cloud. Managed compute, no setup, just an API key and you're off.
We didn't do that.
Your AI agents are already authenticated with your accounts, configured with your MCP servers, set up with your preferences. They run in your filesystem, with your credentials. Extracting that into a cloud environment means re-creating your entire local setup in someone else's infrastructure, paying to store your code and logs on servers you don't control, and accepting latency on every interaction.
The harder path — the right path — is to connect directly to where the work already lives. Seawork does this. The daemon runs locally. The mobile app connects over an end-to-end encrypted relay. Your code, your logs, your credentials: none of it touches our servers.
What supervision actually looks like
There's a common framing in the AI agent space that goes: "set it and forget it." Give the agent a task, come back when it's done.
We think this is mostly wrong — at least for anything that matters.
Agents make decisions constantly. Some are trivial. Some are wrong. The good developers we've watched using these tools don't abdicate judgment, they delegate it with a leash. They check in. They steer. They say "no, not that approach, do it this way" at the right moment.
Seawork is built for that workflow. You can see every tool call your agent makes, read the full conversation, and send a message mid-run to course-correct. You get notified when agents finish, error, or need your approval. Permission requests — the "are you sure?" dialogs that stop Claude Code in its tracks — arrive on your phone so you can unblock them in seconds instead of minutes.
This isn't micromanagement. It's the difference between a good collaborator and an unsupervised process.
Multi-agent is real now
Another thing we've found: serious users are running multiple agents simultaneously. One refactoring a module, one writing tests, one drafting docs. It's genuinely how people work at the frontier.
Managing that from a single terminal window is painful. Seawork gives each agent its own workspace. You can see at a glance what's running, what's waiting, what finished. Switch between them instantly. The interface scales to the workflow.
What we shipped
Seawork 1.0 supports Claude Code and Codex CLI. It runs the native agent harness — your config, your MCP servers, your credentials, untouched. The desktop app ships with a built-in daemon, so there's nothing extra to install on your development machine. The mobile app is on iOS and Android.
Remote access works over an E2E encrypted relay with zero-knowledge architecture. The relay routes encrypted packets; it can't read them.
The gap between where your agents run and where you are is a solvable problem. We solved it.