Integrating AgentVision into your workflow & your agents

Developers use AgentVision in two places: their dev/CI workflow (gate merges on what the UI actually looks like) and their agents (give the agent eyes so it self-corrects). Both on-ramps below.

In your CI / workflow

GitHub Action (installs AgentVision + Chromium and fails the build on a fail verdict):

# .github/workflows/visual.yml
jobs:
  visual:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: amitpatole/agent-vision@v0.6.1
        with:
          source: dist/index.html
          command: check            # check (no key) | analyze | conform | watch
          args: --full-page
      # Intent gate with a vision backend:
      - uses: amitpatole/agent-vision@v0.6.1
        with:
          command: conform
          source: dist/index.html
          backend: anthropic
          expect: |
            must: a "Checkout" button is visible
            should: uses the brand's dark theme
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

pre-commit (this repo ships .pre-commit-hooks.yaml):

# .pre-commit-config.yaml
- repo: https://github.com/amitpatole/agent-vision
  rev: v0.6.1
  hooks:
    - id: agentvision-check
      args: ["dist/index.html", "--quiet"]

Any CI / Makefile / script — just shell out; non-zero exit fails the gate:

agentvision check ./dist/index.html --quiet            # 0 pass/warn · 2 fail · 3 error
agentvision conform ./dist/index.html --expect 'must: shows "Total"' --quiet
agentvision watch http://localhost:3000 --allow-local --no-vision --quiet

--quiet writes only the JSON report to stdout (logs to stderr) — easy to capture and parse. The bundled Dockerfile bakes in Chromium/Tesseract/poppler for a hermetic runner.

In your agents

"Provider-agnostic" means the API surface works anywhere — but an agent only benefits if it's actually told to use the loop. Pick the on-ramp that matches your agent:

Agent / host	How	Recipe
Claude Code	Skill (auto-invokes before "done")	skill/SKILL.md
Cursor	MCP server or project rule	integrations/cursor.rules.md
Aider	/run or --test-cmd	integrations/aider.md
Any MCP host	agentvision-mcp	adapters.md
Anything else	the CLI + a system-prompt contract	integrations/agent-contract.md

The universal contract

Whatever the agent, the behavior you want is:

1. After producing/editing a visual artifact, run agentvision check/analyze <artifact> --quiet (use conform to also grade intent, watch for streaming/over-time behavior).
2. Treat issues as a required to-do list (or use --handoff for next_action/todo); fix the source.
3. Re-run agentvision loop until the verdict is pass.
4. Don't claim the task done without a passing verdict.

Copy integrations/agent-contract.md into your agent's system prompt to enforce it.

Why Claude Code is "first-class"

The Claude Code Skill is the one surface that makes an agent invoke the loop proactively (its description triggers on visual work, before the agent claims completion). Other agents get the same capability but you must wire the trigger yourself (a rule, a test command, or the contract above). MCP is the first-class cross-host path.