Agent-See: SaaS-to-Agent Converter

What This Skill Does

Analyzes a SaaS application (from a URL, OpenAPI spec, or codebase) and generates
a complete agent-native interface — including an MCP server, A2A Agent Card,
OpenAPI spec, and AGENTS.md — with a mathematical proof of coverage and fidelity.

The original software is never modified. The generated interface is a proxy/wrapper.

Prerequisites

• Python 3.12+
• uv package manager (pip install uv)
• For URL analysis: playwright install chromium

Usage

# From a URL (primary use case for SMBs)
uv run agent-see convert https://mybakery.com

# From an OpenAPI spec (highest fidelity, zero hallucination)
uv run agent-see convert ./openapi.json

# Custom output directory
uv run agent-see convert https://example.com -o ./my-output

Output

The tool generates an agent-output/ directory containing:

• mcp_server/ — Deployable MCP server (FastMCP with tools for each capability)
• agent_card.json — A2A Agent Card for discovery
• openapi.yaml — OpenAPI 3.1 specification
• AGENTS.md — Agent capability documentation
• proof/proof.json — Mathematical proof of correctness
• proof/proof_summary.txt — Human-readable proof summary

How It Works

1. Discover: Probes the URL for OpenAPI specs, hidden APIs, and page structure
2. Extract: Pulls capabilities from API specs (deterministic) and browser DOM (LLM-assisted)
3. Map: Builds a capability graph with domains, relationships, and workflows
4. Generate: Creates MCP server, Agent Card, OpenAPI spec, AGENTS.md
5. Verify: Runs coverage proof, fidelity scoring, hallucination check

Proof Guarantees

Every conversion includes a proof.json certifying:

• Coverage score (target: 1.0 = all capabilities covered)
• Fidelity score (target: >= 0.95 = semantics preserved)
• Hallucination check (must pass: no fabricated tools)
• Context efficiency (compression ratio vs raw docs)