ai-vision

ai-vision lets an AI agent read and drive your app through one call tool and named paths (it is not computer vision). It is the engine behind Persephone’s MCP server.

How an AI agent sees Persephone: one call tool, named paths, and boards and web pages that plug their own models in

Core features

• One tool, not hundreds. A self-describing object model replaces a hand-written tool per action.
• Incremental discovery. The agent reads a short overview and opens only the branch the task needs.
• Self-correcting. Members are allow-listed, so a wrong path gets back the valid names and a “Did you mean …?” suggestion.
• Cheap on context. Parts the agent doesn’t look at cost nothing, so small models make few mistakes.
• Pluggable. A board, a web page or another app can publish its own model into the same tree.
• Point at things on screen and keep the agent up to date with an event log.

See it in Persephone

With Persephone’s MCP server enabled, your agent can open pages, edit documents, switch editors, drive the browser and install boards, all through call. Try it by asking your agent to open something in Persephone.

Use it in your app

npm install ai-vision

Source and docs: github.com/andriy-viyatyk/ai-vision.