A token-efficient, provider-agnostic vision layer for reasoning models via MCP.
Token-efficient vision for MCP clients. An MCP server (stdio + Streamable HTTP) exposing a single analyzeimage tool that lets any MCP client see images — describe, OCR, or inspect — through an OpenAI-compatible vision provider.
It exists for one reason: to give a model without native vision (or any client that wants to spend fewer tokens) the ability to understand images. The client sends only an image reference plus a mode; the server resolves the image, sends it to the vision model, and returns compact plain text.
- Two transports — stdio for local clients (Claude Code, Cursor), Streamable HTTP (POST /mcp) for remote / shared / team deployments. - One tool, three modes — describe (summarize), ocr (extract text), inspect (answer a specific question via prompt). - detail: low by default — cheaper and measurably better OCR than high. - Fence-free OCR — every OCR answer is stripped of markdown code fences defensively. - Four image sources — local path, url, base64 (data URL or raw), and resource (resolved against a configured directory). - Hardened — SSRF protection (private/loopback/link-local IPs blocked), DNS resolved before fetch, MIME sniffing (PNG/JPEG/WebP/GIF) + size limits, at most one redirect. - Structured errors — 11-code taxonomy, MCP-safe messages (no keys/URLs leak). - Observability — token/cost estimation and one JSON log line per call, written to stderr (never stdout, which is reserved for the JSON-RPC channel).
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/kasumaputu6633--visbridge-mcp.html)
Local MCP relay for cross-model agent teams. Your Claude, Codex, and other agents talk to each other over Unix sockets. No broker, no network, no accounts.
zouyuanqing/vision-primitives-mcpVision-primitives MCP server for text-only LLMs: describe, locate (coordinates), OCR with bbox, annotate, crop, zoom, automated anomaly scanning, computer use — powered by Xiaomi MiMo V2.5. Single-fil
Alex663028/pulse-agentPulse — Hermes-style self-improving AI agent. Reliability-first rebuild with evaluated skill self-evolution, multi-agent team orchestration, dialectic user modeling, and fully self-hosted default stac
0xwilliamortiz/openclaude-improvedruns anywhere. uses anything
slxca/opencontextA Model Context Protocol server for persistent project-specific AI agent context.
Gowthaman90/mcp-bastionReliability + security proxy for the Model Context Protocol (MCP): self-healing connections, runtime tool-security (rug-pull/poisoning detection), and a compliance-mapped audit trail.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.