prismrelay-mcp

Text-only Agents hit a wall the moment a workflow involves a screenshot, a scanned document, or a chart. PrismRelay MCP addresses this gap head-on: without modifying the underlying text model, it attaches a callable external vision capability via a local stdio MCP server, turning image understanding, generation, and editing into standard tool calls within the Agent workstream.

Three MCP tools form the core. prismrelay_understand_image inspects image pixels to answer questions about scenes, objects, visible text, and diagrams. prismrelay_generate_image and prismrelay_edit_image provide experimental generation and editing flows with optional references, dimension control, and manual review checkpoints. Visual reasoning is delegated to the Agnes AI backend, while the Agent remains the primary reasoner responsible for initiating calls and integrating results.

Typical scenarios include diagnosing UI screenshots and error dialogs, extracting headings and dates from document images, interpreting chart trends and legends, comparing two product or interface screenshots, and verifying whether a poster contains a specified element. It is particularly useful for developers who have built text-only Agent pipelines in DeepSeek Harness, Claude Code, or Codex and need to close the vision gap. PrismRelay is not a replacement for a native multimodal model; visual accuracy ultimately depends on the Agnes AI service.