Frequently asked questions

On the files themselves. Your app opens an outbound connection, and file operations travel back down it to run on your machine — so an agent reads the file you have open, writes the change into your working copy, and runs the build where the project lives. There's no upload step and no snapshot to keep in sync.

For a full account of what is stored where, see the security page.

Any OpenAI-compatible endpoint. That covers hosted providers such as Groq and OpenRouter, a gateway you run yourself, and models running on your own hardware. You can set a different model per project, and change it whenever you like.

You bring your own model access, and we never see your key. We don't resell inference, which means we can't mark it up and you're never locked into a house model.

Yes. Point WorkByte at the internal endpoint. The inference request travels over your own bridge to that endpoint, so the model stays inside your network and never needs to be publicly reachable.

No. All connections are outbound from your machine.

An assistant completes what you're typing. WorkByte takes a goal, plans it into milestones, splits it across specialist agents, executes them, and reports back with evidence of what was done. It has your whole project in view — files, notes, a semantic index and a code graph — not just the buffer you have open.

Uploaded files are a snapshot. A WorkByte project is live: the agents read your current files, write changes back, run builds where the code lives, and keep a project-scoped index that updates as the code does.

Yes — clone, branch, commit, push, open a pull request. Maven builds run in a container against the real project root. (Gradle is not supported through that path yet.)

A run has to produce evidence — files written, commands run, output captured — and the platform checks for it before reporting success. This is a deliberate feature, added because "it's done" without artifacts is the characteristic failure of agent systems.

You get a report saying where it stopped and why, along with the progress log of everything that did complete — so a failed run still leaves you the work it finished.

Everything the product can do. The tier decides where the scratch work runs — on the free tier, scripts and media processing use your machine; on a paid tier, our cloud. Your code is on your own machine either way.

Windows desktop today. The codebase is cross-platform; other builds aren't distributed yet.

Connect any MCP server and its tools become available to your agents.

Yes, from the UI: name, instructions, model, tools. Agents are data rather than code in this platform, so the ones you create are the same kind of thing as the ones that ship with it.

We don't train models. Your prompts go to the provider you chose — their policy is the one that applies, so check it. If that's a problem, use a model on your own hardware.

We don't hold a security certification and we don't make a compliance claim. What we can tell you is exactly where your data goes — that's the security page. If you have a hard requirement, get in touch before you rely on us for it.