AgentVision — Eyes for AI Agents 👁️

Problem: AI coding agents are blind — they write a UI, chart, SVG, PDF or image and never see the result, shipping breakage they can't perceive. Result: AgentVision gives them eyes — render → see → report → fix — so the agent self-corrects before it claims done.

pip install "agentvision[render]"
playwright install chromium      # see `agentvision doctor` if Chromium won't launch
agentvision demo                 # no API key required

agentvision demo renders a deliberately broken page, prints a FAIL report (overflow + low-contrast + a 404 image — all DOM/CV-grounded, no LLM key), then loops against the fixed version and prints "what changed: 3 issues resolved → PASS."

What it does

Capability	What you get
See & report	A machine verdict (pass/warn/fail) + coordinate-grounded issues — DOM geometry, computed-style WCAG contrast, clipped/truncated text (incl. SVG chart labels & PPTX text boxes), OCR/typos, broken-image & console/4xx capture.
Match intent	Grade a render against a brief / checklist / reference — PASS means "it's what I set out to build," not just "defect-free."
Full-coverage vision	Large artifacts get a downscaled overview plus full-res tiles — pixel-based & source-agnostic (HTML, image, PDF, canvas).
Streaming / temporal	watch verifies behavior over time — playback, loading, liveness — not just a single glance.
Eyes → brain handoff	A distilled {verdict, next_action, todo, open_questions} signal any agent/brain acts on.

Where to go next

• :material-rocket-launch: Quickstart — install, system deps, first run.
• :material-hand-wave: Try it yourself — a key-free, copy-paste FAIL→PASS in minutes.
• :material-school: 5-minute tutorial — check → analyze → conform → loop.
• :material-lightbulb-on: Use cases · :material-play-box: Real-world scenarios — runnable demos, real output.
• :material-sync: The loop — render → perceive → report → fix → diff.
• :material-target: Conformance — grade against intent.
• :material-server-network: Swarms & scaling — eyes as a service for a fleet of agents.
• :material-brain: Handoff — wire perception into your reasoning/memory.
• :material-cog: Workflows & agents — GitHub Action, pre-commit, MCP, the agent contract.
• :material-console: CLI reference · :material-tune: Configuration · :material-language-python: Python API
• :material-book-open-variant: Recipes · :material-help-circle: FAQ

Many agents, one set of eyes

One agent with eyes self-corrects. A swarm of agents sharing one set of eyes is the real prize — dozens of workers each rendering UIs, charts, decks or PDFs, every output graded against the same contract before it counts as done. Run the eyes as a horizontally-scaled service (or embed the library per worker); single-shot grading scales with zero coordination, and because every worker returns the same agentsensory Report, a coordinator or a brain like Verel aggregates all the verdicts on one bus. See Swarms & scaling.

Eyes & brain

AgentVision is the eyes. It pairs with Verel, the brain — an agent framework where nothing is "done" until a grader returns a verdict. The eyes perceive and grade intent; the brain decides and compounds only verified work into memory; then the eyes look again.

Install: pip install "agentvision[all]" · Source: GitHub · Package: PyPI · License: MIT.