Adapters

One core engine, several faces. All adapters wrap the same agentvision.core functions and the same Report model.

Library

import asyncio
from agentvision import load_settings
from agentvision.core import analyze, check, render, compute_diff, contact_sheet
from agentvision.core.loop import LoopSession

report = asyncio.run(analyze("index.html", settings=load_settings(vision_backend="local")))

CLI

agentvision demo | render | analyze | check | diff | ocr | loop | sheet
              | baseline | regress | doctor | serve

All commands accept --json and exit non-zero on a FAIL verdict.

MCP server

agentvision-mcp        # stdio

Register it with an MCP host (Claude Desktop / Cursor):

{
  "mcpServers": {
    "agentvision": { "command": "agentvision-mcp", "env": { "ANTHROPIC_API_KEY": "sk-ant-…" } }
  }
}

Tools: analyze_artifact, check_artifact, render_artifact, contact_sheet, visual_diff, ocr_artifact, start_loop, loop_iterate, manage_baseline, doctor. Image-returning tools downscale before encoding so they don't overflow stdio/token budgets. Loop sessions persist in-process so loop_iterate continues a start_loop.

REST service

agentvision-serve --host 0.0.0.0 --port 8000

Endpoints: POST /analyze /check /loop /loop/{id}/iterate /sheet /baseline, GET /baseline/{name} /artifacts/{id} /healthz.

Security: API keys are read server-side only and are never accepted in requests. Per-request backend is limited to the server allowlist (AGENTVISION_REST_ENABLED_BACKENDS, default ["local"]).

Multi-worker caveat: loop sessions are kept in-process. Behind multiple workers you need sticky sessions or a shared session store; for single-shot analyze/check there is no such constraint.