@devrobotlabs/visionapi-mcp

MCP server for the Vision API. Point Claude Code, Claude Desktop, Cursor or any other MCP host at a folder of scans and ask for the invoices — no integration to write, no API key in generated code, no contract paraphrased from memory.

You: pull the totals out of every invoice in ~/inbox and put them in a CSV

Claude: [vision_analyze × 7]
        Done — 7 invoices, 14 credits. Three had no PO number; I left those cells empty.

Install

Nothing to install. Add it to your host's config and it runs via npx.

Claude Code — claude mcp add visionapi --env VISION_API_KEY=sk_live_... -- npx -y @devrobotlabs/visionapi-mcp ~/inbox

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "visionapi": {
      "command": "npx",
      "args": ["-y", "@devrobotlabs/visionapi-mcp", "/Users/me/inbox"],
      "env": { "VISION_API_KEY": "sk_live_..." }
    }
  }
}

Cursor — .cursor/mcp.json, same shape:

{
  "mcpServers": {
    "visionapi": {
      "command": "npx",
      "args": ["-y", "@devrobotlabs/visionapi-mcp", "."],
      "env": { "VISION_API_KEY": "sk_live_..." }
    }
  }
}

VS Code — .vscode/mcp.json:

{
  "servers": {
    "visionapi": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@devrobotlabs/visionapi-mcp", "${workspaceFolder}"],
      "env": { "VISION_API_KEY": "sk_live_..." }
    }
  }
}

Get a key at app.visionapi.io/dashboard/keys. New accounts get 50 credits, no card.

The host config blocks live in this README rather than in an examples/ directory — a departure from the nine client libraries, and a deliberate one. A snippet that exists only in a README cannot drift from a runnable script nobody runs.

Related MCP server: receiptconverter-mcp

Which directories it can read

Every positional argument is a directory the server may read files from. With none given, the process working directory is the only root — which is the safe default, because MCP hosts launch a stdio server with the project directory as its cwd.

This matters more than it may look. The server holds a live spending credential and runs with your filesystem permissions, so it can read anything your shell can. Asking it for the raw text of ~/Documents/passport.jpg is a working way to get that document's contents into a model's context and into whatever transcript your host keeps. The allowlist is what stops a confused or manipulated agent doing that by accident.

Paths are resolved with realpath on both sides before being compared, so a symlink inside an allowed directory cannot point out of it.

--allow-any-path turns the allowlist off entirely. It warns on stderr at startup and you should have a reason.

Tools

The three that spend credits are annotated readOnlyHint: false, so a host that auto-approves read-only tools will still stop and ask before one of them runs.

The catalogue is also exposed as resources — visionapi://presets and visionapi://presets/{name} — for hosts that prefer them. Tools are the primary surface, because resource support is uneven across hosts.

Output

Responses are rendered for a model to read, not dumped as JSON. A 37-field invoice preset over a document that fills twelve of them comes back as a table plus one Not found in this document (25): … line, rather than as twenty-five repetitions of {"value":null,"confidence":"low"} — three to four times smaller, and easier to act on.

Nothing is lost in the process. Every tool takes a format:

• markdown (default) — the rendering above.

• compact_json — the same information as data, with _not_found and _low_confidence arrays. For when the agent will parse rather than read.

• json — the API response verbatim. What to reach for when writing real HTTP code against the contract.

Confidence is printed only when it is not high, so (mid) and (low) stand out and the common case costs nothing to read.

Long documents

Leave mode at auto. The API kills a synchronous request at 60 seconds; the server then re-submits it to the queue and polls, reporting progress to your host as it goes. You are charged once, because the timed-out attempt released its reservation in full.

Pass mode: "async" up front for anything over roughly ten pages, and pages: "1-5" to sample a long document cheaply — you are charged for selected pages only.

Costs and failures

Failures cost nothing. Every non-2xx releases the credit reservation in full, so a failed call is safe to correct and repeat and there is no cleanup to do. The tool descriptions say so, which is why an agent using this server behaves sensibly after an error instead of either giving up or retrying something that cannot work.

Two errors carry advice that is worth knowing yourself:

Environment

A missing key does not stop the server starting: it warns on stderr, tools/list still works, and the first billable call returns a message naming the fix. A server that refuses to start tells the user only that something is broken.

Development

npm install
npm run typecheck
npm test          # 37 offline tests — no key, no network
npx @modelcontextprotocol/inspector node ./dist/cli.js ~/some/dir

npm install --no-save ../node to test against a local build of the client. Not npm install ../node — that rewrites package.json to "file:../node", and that manifest is what gets published.

Links

• Documentation

• Integration reference for coding agents

• Preset catalogue

• Pricing

MIT licensed.