Local ocr vlm

Skill batteryshark/skill-tap/skills/utilities/local-ocr-vlm

Portable, reusable skills for agents and everyday workflows. Clank responsibly.

Install

npx -y skills add batteryshark/skill-tap --skill local-ocr-vlm

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

27 days oldThe repository was created 27 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Run OCR on local images through an OpenAI-compatible local vision endpoint (LM Studio, vLLM/vmlx). Use when you need offline, on-machine document OCR into markdown, including batch directory processing, without sending images to a hosted service.

SKILL.md

2.2 KB, 500 tokens by cl100k_base, as published. Nobody here has run it

Local OCR VLM

Send image(s) to a locally hosted OpenAI-compatible vision endpoint and capture the text. Defaults to http://127.0.0.1:1234/v1. The tool is a single, zero-dependency Node script; no Python or SDK is required.

When to use

OCR screenshots, scans, or document images on-machine.
Batch a folder of images into per-image markdown files.
Keep OCR offline using LM Studio or a vLLM/vmlx-compatible API.

Prerequisites

Node.js 18+ (uses the built-in global fetch; no packages to install).
A local OpenAI-compatible server exposing a /v1 API with a vision-capable model loaded. Qwen 3.5 9B is preferred; Qwen 3.5 27B also works.

Commands

Run the portable entry point:

bin/local-ocr-vlm invoice.png
bin/local-ocr-vlm --input-dir doc_input --output-dir doc_output --max-workers 4
bin/local-ocr-vlm --input-dir doc_input --model "qwen3.5-vl-9b"

Run bin/local-ocr-vlm --help for every flag. See references/reference.md for the endpoints called, model auto-selection, thinking-mode handling, and exit codes.

Workflow

Confirm a local endpoint is serving /v1 with a vision model loaded.
Point the tool at one or more images and/or an --input-dir. With no --output-dir, results print to stdout; with one, each image becomes <stem>.md.
Let the tool auto-select a model, or pass --model when the endpoint hosts several.
Tune --max-workers for throughput; use 1 for strictly sequential processing.

Rules

Never point the tool at a remote/hosted endpoint when the intent is to keep images on-machine; the default base URL is loopback.
The tool reads images and writes only markdown output; it does not modify input images.
Treat an unreachable endpoint or a per-image request failure as an explicit error. A run exits non-zero if any image fails.

Use agents/ocr-operator.md when a subagent should choose flags and verify the endpoint before running a batch.