The MCP connector for software with no API. Your AI delegates. Fleet does the clicking.
Claude, ChatGPT and the agents you build are good with code, APIs and the web. What's left is the work inside desktop software with no API: Sage 50, Access, old ERPs. Add Fleet AI as a connector, sign in and pick the computers it may use. Your AI hands that work off, waits for it, and reads back exactly what happened.
Step by step: Claude · ChatGPT. Connecting Claude, ChatGPT or your own agent is free, and so is driving a computer yourself over MCP. Runs Fleet does for your agent are billed like any other run. Pricing →
# Add Fleet once. No token to paste.
claude mcp add --transport http fleet \
https://app.tryfleet.ai/mcp
# Type /mcp, sign in, pick your computers.
# Then just ask
> Reconcile last month's bank feed in Sage 50
on the accounting PC, then send me the exceptions.Three calls: hand off, wait, read back.
- run_workflow · run_task1
Hand off the job
Your agent picks a computer and starts a saved workflow, or sends one complete instruction: which app, which company file, which period, what counts as done.
- wait_for_task2
Wait while Fleet works
Fleet's own agent does the work on the Windows computer, under Fleet's rules: plan approval, checkpoints and a budget. Your agent checks in until it's done or needs a person.
- get_run_activity · get_run_artifact3
Read what happened
A short digest of the run (plan, actions, notes, errors), the screenshot of any step, and the files it saved. Your agent quotes the numbers from the files, not from memory.
The full set also covers listing computers, stopping runs, saving workflows and proposing schedules.See the tool reference →
Your agent could take screenshots and click. It shouldn't.
General computer use works for a one-off task with someone watching. For work that has to be cheap, come out the same way every time and be checked afterwards, hand it to something built for that.Why Fleet AI →
Cheaper than clicking itself
Driving a desktop app by screenshot fills your agent's context with images and costs a model call per click. Fleet reads the app's own data instead, at more than 10x lower cost on our Access tests. Your agent only sees a task id and a result.
The same way every time
Saved workflows are the tested way to do a job, and Fleet's code does the matching and checks every change against the app. Your agent can save a new workflow when it finds a good way to do something.
A record it can learn from
When a run fails, your agent doesn't have to guess why. The digest keeps the problems and the outcome, and the screenshots show exactly what was on screen.
Three ways to give an AI your desktop software.
A server that only exposes click and type tools, a vendor that wraps one legacy app in an API, or a connector that takes the whole job.
| Compared on | A click-tool MCP server | A legacy-app API vendor | Fleet AI |
|---|---|---|---|
| Who does the clicking | Your assistant, one screenshot and one click at a time | The vendor's scripts, behind an API your developer calls | Fleet's own engine, on your Windows computer. Your assistant only hands off the job |
| Who approves | Whatever your assistant's app asks you to confirm | Whatever your developer builds in | A person in Fleet, by default: plans and questions wait |
| Cost per run | High: the assistant pays for every screenshot and every click | Set by the vendor, per app it supports | Low: more than 10x lower than a screen-only AI agent on our Access reconciliation* |
| Same result every run | Not guaranteed: each click is decided afresh | Yes, for the actions the vendor has built | Saved workflows; code does the matching and checks every change |
| Run record | A chat transcript | API logs | The plan, every action, a screenshot of each step and the files it saved |
| Setup | Add the server, then guide every step in the chat | A developer integrates each app | Add the connector, sign in, pick your computers |
* Single runs measured on Fleet AI’s production platform, August–October 2026. The screen-only agent and Fleet AI used different AI models.
Your agent gets a desktop, not the keys to the business.
The connection is the boundary
When you sign in, you pick the computers the app may use, and whether it may hand off tasks, drive a computer itself, or both. It sees only the runs it started. Pause or disconnect it in Fleet and it stops at once.
People answer approvals
By default a plan or a question waits for a person in Fleet. Only turn on answering for an app you trust to approve its own runs.
A monthly cap per app
Runs count against your Fleet credits and the connection's own monthly cap. When either runs out, new runs are refused.
Schedules start switched off
Your agent can propose a schedule for a workflow. It is created disabled, and a person turns it on in Fleet.
No passwords over MCP
Fleet types saved logins from its encrypted vaults itself. They never pass through your agent.
Every call on the record
Runs your agent starts show up in Fleet like any other, live view included. Starting, stopping or answering a run is written to the audit log.
Any app that speaks MCP.
You sign in to Fleet and pick the computers the app may use: no token to paste. Apps that can't sign in this way connect over Streamable HTTP with a connection token. There's a plain-text copy of the guide for agents attryfleet.ai/docs/mcp.md.
Show us the desktop work your agent can't reach.
Tell us the task and we’ll show you Fleet on it. Or create an account and start with $10 of credits.