# F12: Agent Vision & Browser Automation — PRD — HC v1.3.3
> Source: https://ideas.asapai.net/f12-agent-vision-prd
> Type: prd | ID: f12-agent-vision-prd

---
## hc-metadata
{
  "hc_version": "1.3.3",
  "hc_type": "prd",
  "artifact_id": "f12-agent-vision-prd",
  "title": "F12: Agent Vision & Browser Automation — PRD",
  "created": "2026-02-25",
  "updated": "2026-02-25",
  "author": "jason",
  "status": "draft",
  "tags": [
    "prd",
    "feature",
    "vision",
    "browser",
    "playwright",
    "forge",
    "agent"
  ],
  "summary": "PRD for giving Claude Code and Forge agents the ability to see web pages, take screenshots, click elements, and visually verify deployed work. Playwright-based with MCP server wrapping for universal agent access.",
  "registry_stack": [
    {
      "level": "meta",
      "url": "https://ideas.asapai.net/meta-registry"
    },
    {
      "level": "project",
      "url": "https://ideas.asapai.net/meeting-intelligence-registry-athio"
    }
  ],
  "related_pages": [
    {
      "url": "https://ideas.asapai.net/nowpage-issue-tracker-2026-02-20",
      "relationship": "parent-prd"
    },
    {
      "url": "https://ideas.asapai.net/hc-page-agent-architecture",
      "relationship": "related-architecture"
    },
    {
      "url": "https://ideas.asapai.net/phase-1-internal-execution-v2",
      "relationship": "sprint-context"
    }
  ]
}

---
## hc-instructions
This is a Product Requirements Document (PRD) for Feature F12: Agent Vision & Browser Automation.
    It defines the capability for AI agents (Claude Code sessions, Forge VPS agents) to visually
    interact with web pages — taking screenshots, clicking elements, filling forms, and verifying
    deployed work. The implementation lives in the Forge repo, exposed as an MCP server that any
    agent can consume. Read this to understand the architecture, phases, and acceptance criteria.

---
## hc-context-public
{
      "feature_id": "F12",
      "feature_name": "Agent Vision & Browser Automation",
      "priority": "high",
      "effort": "medium",
      "owner": "jason",
      "implementation_repo": "forge",
      "consumer_repos": ["folio-saas", "outreach-agents"],
      "phases": [
        {
          "phase": 1,
          "name": "Playwright MCP Server",
          "status": "not-started",
          "deliverables": ["mcp-browser-server", "screenshot tool", "navigate tool", "auth helper"]
        },
        {
          "phase": 2,
          "name": "Claude Code Integration",
          "status": "not-started",
          "deliverables": ["MCP config for claude code", "post-publish verification", "wizard testing"]
        },
        {
          "phase": 3,
          "name": "Forge VPS Headless",
          "status": "not-started",
          "deliverables": ["headless chromium on VPS", "systemd service", "agent tool bindings"]
        },
        {
          "phase": 4,
          "name": "Antigravity Evaluation",
          "status": "not-started",
          "deliverables": ["evaluate vs raw playwright", "decision doc", "migrate or stay"]
        }
      ],
      "key_decisions": [
        "Playwright as base layer (proven, headless, auth-capable)",
        "MCP server wrapping (universal agent access)",
        "Implementation in Forge repo (agent infrastructure, not app code)",
        "folio-saas consumes via MCP (post-publish visual verification)"
      ]
    }