Local MCP tools for agentic AI

Give any AI hands.

A free, open-source, Windows-native MCP server. Browser control, native desktop automation, vision, and OCR - so the model can operate the interface and verify what actually happened. Bring your own AI.

Works with Claude Code, Codex, Grok Build - anything that speaks stdio MCP. Signed installer.

ai-hands - session ACT, THEN VERIFY
> you: open the report, pull the March total, prove it browser_navigate reports.example ok browser_a11y_snapshot read page structure ok uia_click "March" ok read_screen_text expect a total $48,210 vision_screenshot proof_march.png saved > done. read it back off the screen to confirm.

Illustrative session (placeholder) - a real screen recording replaces this before launch.

Route by reality

Some work has an endpoint. The rest has buttons.

Use a documented or authorized API when one exists. When the work still lives behind a screen, AI-Hands gives the model a way to operate the interface - and then check the real outcome before it says "done."

01User intent

Describe the result you want, in plain language.

02A - APIAuthorized API

When a documented or permitted endpoint is available.

02B - UIUser interface

When the task lives on the screen.

03Verify

Check the real outcome on screen before reporting success.

No account setup, no bypassed authentication, no undisclosed private endpoints. AI-Hands does not grant permission - use only interfaces and APIs you are authorized to access.

Capability layers

The right surface for the screen.

AI-Hands is an MCP capability surface. Your chosen compatible client supplies the model; AI-Hands supplies the routes into the browser, Windows, and the visible screen.

01 / BROWSER

Read the page structure.

Navigate pages and work with tabs, forms, downloads, and semantic controls.

02 / WINDOWS

Reach native controls.

Use apps, menus, dialogs, and Windows interface elements outside the browser.

03 / VISION

Look when structure stops.

Screenshots and visual analysis for interfaces with no useful semantic route.

04 / VERIFY

Check before "done."

Inspect visible state and screenshots before the model reports the result happened.

How it works

A thin layer between intent and interface.

Three steps from install to a verified result.

01 CONNECT

Add AI-Hands

Add AI-Hands to a compatible MCP client on Windows. One signed installer.

02 ASK

Describe the result

Say what you want done, in plain language - no scripts, no selectors.

03 ACT + CHECK

Operate and verify

The model uses the available interface route, completes the task, and inspects the outcome.

Lineage

An AI built the tool that built everything else.

AI-Hands wasn't written the usual way. The first thing built was a coding server - the contractor. That contractor built the next tool, and the next. By the time AI-Hands existed, the toolchain was building itself: an AI, given hands, extending its own reach one server at a time.

What came from a solo founder with no programming background was the intent. What came from the tools was the execution - repo after repo, signed release after signed release, the hands doing the work once they existed.

[[STORY PLACEHOLDER - Joseph to confirm the real detail: last December, hands set up a Christmas song for the family. This is a staged draft; the true anecdote goes here, verbatim, before publish. Not written from memory - awaiting the sourced detail.]]

- how AI-Hands got here

Install

Free, open source, signed.

Download the signed installer, add it to your MCP client, and ask for a result. Bring your own AI - AI-Hands never asks for a model key.

Add to your MCP client (example)
{
  "mcpServers": {
    "hands": {
      "command": "C:\\Program Files\\AIWander\\hands.exe",
      "args": [],
      "env": {}
    }
  }
}
Local-firstRuns on your machine. Your data, files, and screen never transit our servers.
SignedEvery binary and installer is Authenticode-signed. No unsigned downloads.
Bring your own AIWorks with any stdio MCP client. We never bill for tokens or hold a model key.
120+ toolsBrowser, Windows UI automation, vision, and OCR - one server.