mac-local-vision CLI
A single-binary macOS tool, macvis: OCR, word coordinates, barcodes and face grouping on-device, with no cloud vision tokens.
- 63 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The tool reads local images on-device and the `serve` mode opens HTTP access.
Install
- Works in
Manual install
brew install junmo-kim/tap/macvisThen `macvis doctor`; MCP connects with `macvis mcp`.
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The tool reads local images on-device and the `serve` mode opens HTTP access.
This is third-party code. Review the repository files before installing.
What it does
`macvis` is written in Swift and relies on the Vision and FoundationModels frameworks built into macOS, so it ships as a single binary of about 550 KB with no Node or Python. `ocr` extracts page text, `find` returns the exact center of a word for clicks or E2E test assertions, `barcode` reads QR codes and barcodes, `make-qr` draws a QR code and `rectify-document` straightens a photographed document. There is also structured document parsing with tables and lists, classification against a taxonomy of 1303 labels, sorting photos by face and an `ask` command for questions about an image via Apple Foundation Models on macOS 27. An agent connects `macvis mcp` over stdio, the HTTP mode `macvis serve` is for remote nodes and the repository holds a Claude Code skill.
Who it is for. Mac developers who ask an agent to read screenshots and scans and do not want to pay for cloud vision or send pictures out.
Good fit when
- A UI test must find a word on screen and click its center.
- A scan or photo of a document must become text and tables.
- A QR code or barcode on an image has to be read.
- Family photos are sorted by person and the pictures must not leave the computer.
Not a fit when
- Work runs on Windows or Linux: the tool relies on macOS frameworks.
- Recognition of handwriting or complex diagrams by a GPT-class model is needed.
- The Mac has an Intel processor or macOS older than 26.
Example request
Find the button with the text Sign in on the screenshot and return its center coordinates, then decode the QR code on the second image.Limitations
An Apple Silicon Mac with macOS 26 or newer is required. The `ask` command works only on macOS 27 with Apple Intelligence turned on, and building from source needs Xcode 27. The binary from a direct download is ad-hoc signed and not notarized, and the system silently kills it unless the quarantine attribute is cleared. Latencies in the README were measured on a few Macs and depend on hardware. The `serve` mode opens HTTP access and you have to protect it yourself.
How to disable. Remove the binary with `brew uninstall macvis` and the `mac-vision` entry from the MCP configuration. Delete the `~/.claude/skills/macvis` link by hand.
FAQ
How do I install it?
Run `brew install junmo-kim/tap/macvis` or `mise use -g github:junmo-kim/mac-local-vision`, then `macvis doctor` shows what is available.
How do I connect it to an agent?
Through MCP: `macvis mcp` works over stdio. Claude Code also has a skill in `skills/macvis`.
Do images go to the cloud?
No, processing runs on the device through Vision and FoundationModels and no cloud tokens are spent.
What does the ask command need?
macOS 27 and Apple Intelligence turned on. On macOS 26 the other Vision commands work but `ask` is unavailable.
What is serve mode for?
It runs an HTTP MCP server for remote and non-Mac nodes that call a Mac for recognition.
Discussion
No questions or reviews yet.
Related
CLI toolsA local MCP server for editing and analyzing images: the agent crops frames, finds objects and reads text from photos
Gives text-only models vision: images go to GLM and Agnes, OCR to Baidu and Windows, PDFs to MinerU, the answer comes back as JSON
Reelpack breaks a local video into transcript, cuts, on-screen text and music beats so an agent reads it without watching.
A local MCP server for voice work on macOS: voice cloning, speech synthesis and audiobooks from PDF on an agent's command
The eyes CLI and skill for agents: a video index, search by words and on-screen text, close-up frames and editing metrics
VidMoat MCP
Vidmoat MCP
Remote MCP server of a cloud video editor: an agent imports footage, edits the timeline, checks frames and renders MP4
README in short
The README covers features, size and speed, requirements, installation through Homebrew and mise, commands, MCP and skill setup and the internal service design.