| name | vision-skills |
| description | Use when the user asks you to analyze, describe, read, or extract text from an image, screenshot, photo, document, or UI screenshot. Also use when they mention 'vision', 'see this image', 'OCR', 'read this picture', 'what's in this image', or 'analyze this screenshot'. Do NOT trigger on unrelated file operations (code editing, system commands, etc.). |
Vision Skills
Use the registered vision-skills MCP tools directly when the user provides or refers to an image. Do not start another server or manually send JSON-RPC through the shell. If the tools are unavailable, state that the MCP integration is not connected, ask the user to run setup-integrations.bat on Windows or setup-integrations.sh on macOS/Linux from the repository root, and have them restart the client. The setup installs dependencies, builds locally, and configures the client to run the clone's absolute MCP entry path.
Tools
| Tool | Inputs | Result |
|---|
analyze | image, optional auto or fixed mode, optional depth | Bounded structured vision response |
analyze_text | image, optional auto or fixed mode | Bounded plain-text summary |
health | None | Provider readiness |
clipboard | None; Windows only | Official MCP PNG image content from the system clipboard |
image accepts a file path, HTTP(S) URL, base64 string, or image data URI. When the user asks about an image currently on the Windows clipboard, call clipboard and analyze the returned MCP image content directly. If the client cannot expose that content to the model, the client must mediate a subsequent analysis call; the tool does not return a data URI.
Mode Policy
MCP analysis accepts auto and fixed modes. Omitting mode inherits the configured SDK default, which is auto unless changed.
| Mode | Exact policy |
|---|
basic | OCR only; no combined structured fields, semantic relationships, or reasoner |
standard | OCR + object + UI detection and combined tables/regions/layout/code fields; no semantic relationships or reasoner |
advanced | Standard plus semantic relationships; no reasoner |
full | Advanced plus reasoner |
Choose basic for quick text extraction, standard for normal screenshots/documents/photos, advanced when relationships matter, and full only when a reasoned interpretation is needed. Use depth: deep for dense or long images with small text; it creates additional tiled provider calls. Otherwise keep fast.
Reading Results
- Final bounding boxes are pixel objects:
{ x1, y1, x2, y2 }.
entity.confidence: null means the provider did not report confidence; do not invent a score.
- Top-level
confidence is an aggregate provider signal, not calibrated end-to-end accuracy.
- Use
provenance to identify the requested mode, selection reason, provider names, and cache hits.
- Treat
errors, warnings, reasonerOutput.openQuestions, and output truncation as material caveats.
MCP output is wrapped with truncation metadata. Read untruncated structured data from data. If truncation.truncated is true, json contains a bounded partial serialization and must not be presented as complete analysis.
Cancellation and progress are protocol features: let the client cancel long calls, and expect progress only when it supplied a progress token.
URL Safety
Image URL checks block redirects, explicit ports, credentials, and non-public resolved addresses, but DNS is not pinned to the connection. For untrusted URLs, acknowledge the DNS rebinding/time-of-check caveat rather than describing URL fetching as a complete SSRF sandbox.