Introducing the Perceptron MCP Server
Today we're introducing the Perceptron Vision MCP Server, which gives any AI agent enhanced understanding of the digital and physical world – without forcing developers to route every task through expensive frontier models.
Perceptron is an AI research lab building the foundations of physical intelligence. Our Isaac-0.2-2B and Isaac-0.2-1B models are best-in-class, powering systems that can see the world, reason through complexity, and act.
Perceptron is an AI research lab building the foundations of physical intelligence. Our Isaac-0.2-2B and Isaac-0.2-1B models are best-in-class, powering systems that can see the world, reason through complexity, and act.
Powered by our vision language models, our open-source server enables agents to see, reason, and act on local and web files alike.
At its core is our Isaac model family, which delivers best-in-class VQA, OCR, object detection, and grounding at a fraction of the cost of general-purpose multimodal models. Instead of spending large token budgets on image-heavy workflows, developers can route vision tasks through Perceptron Vision MCP tools and get back structured output. The result is a more token-efficient stack that turns large data stores into instantly queryable assets.
You can install the Perceptron Vision MCP Server in one line for Claude Code:
API keys available on Perceptron Platform.
What our developers are building
We’ve already seen developers build tools on top of the Perceptron Vision MCP Server:
Desktop organization. Point an agent at a cluttered folder and let it sort hundreds of files into labeled subfolders by content: screenshots, receipts, memes, documents, photos. No manual tagging required.
Photo curation. Give an agent a camera roll or event album and ask it to surface the best shots, flag duplicates, or pull every photo of a specific setting.
Expense reporting. Drop a folder of receipt photos. The agent reads each one, extracts vendor, date, and total, then compiles everything into a structured spreadsheet ready to submit.
What the server can do
The server exposes five capabilities as standard MCP tool calls:
question()
Ask anything about an image and get structured responses. "What error is shown in this screenshot?" "How many items are on this shelf?" "Does this UI match the design spec?"
question()
Ask anything about an image and get structured responses. "What error is shown in this screenshot?" "How many items are on this shelf?" "Does this UI match the design spec?"
caption()
Generate concise or detailed image descriptions. "Describe this product photo for our catalog." "Summarize what's happening in this security camera frame."
caption()
Generate concise or detailed image descriptions. "Describe this product photo for our catalog." "Summarize what's happening in this security camera frame."
caption()
Generate concise or detailed image descriptions. "Describe this product photo for our catalog." "Summarize what's happening in this security camera frame."
ocr()
Extract text from images in plain, markdown, or HTML format. "Extract all line items from this receipt." "Pull the table data from this PDF." "What does the sign in this photo say?"
ocr()
Extract text from images in plain, markdown, or HTML format. "Extract all line items from this receipt." "Pull the table data from this PDF." "What does the sign in this photo say?"
ocr()
Extract text from images in plain, markdown, or HTML format. "Extract all line items from this receipt." "Pull the table data from this PDF." "What does the sign in this photo say?"
detect()
Locate and identify objects with optional class filtering. "Find all buttons in this UI mockup." "Locate every person in this image." "How many boxes are on this pallet?"
detect()
Locate and identify objects with optional class filtering. "Find all buttons in this UI mockup." "Locate every person in this image." "How many boxes are on this pallet?"
detect()
Locate and identify objects with optional class filtering. "Find all buttons in this UI mockup." "Locate every person in this image." "How many boxes are on this pallet?"
list_models()
List available Perceptron models and their capabilities.
list_models()
List available Perceptron models and their capabilities.
list_models()
List available Perceptron models and their capabilities.
All tools accept URLs or local file paths. Together, they enable frontier image analysis, from traditional computer vision and OCR to advanced understanding and niche object detection.
How it works
The @perceptron-ai/mcp-server npm package runs locally on your machine and connects to the Perceptron Vision MCP Server. When you reference a local file, the server reads it from disk, efficiently uploads and passes the resulting URL to the Perceptron Vision MCP Server. When you reference a web file, the resulting URL simply passes through to the server. In both cases, structured results come back to your agent through the standard MCP tool response.
Getting Started
One-click Install
Alternative setup
Claude Code
claude mcp add perceptron -e PERCEPTRON_API_KEY=YOUR_API_KEY -- npx -y @perceptron-ai/mcp-server@latest Codex
codex mcp add perceptron --env PERCEPTRON_API_KEY=YOUR_API_KEY -- npx -y @perceptron-ai/mcp-server@latest The package downloads from npm when your client starts, always up to date. Get your API key on the Perceptron Platform. See our MCP guide for full setup instructions.
Join our community
If you're building with vision and MCP, we’d like to hear from you! Join our Discord community or drop a note to support@perceptron.inc
