Legal Analysis

Build MCP AI Agents for Legal — Dr. Maryam Miradi's 5-step framework, fully implemented.

AI-powered legal document analysis pipeline: parse → understand → generate → audit → route. One MCP server, reusable from LangGraph, CrewAI, and Google ADK.

Legal PDF
    │
    ▼
Step 1: Scope/Objective
    LegalAgentState initialised (confidence=0, escalation=True, validated=False)
    Document type check: accept contracts/agreements/NDAs/leases/employment
    Reject: invoices, scripts, unknown → escalate immediately
    │
    ▼
Step 2: Ground/Observe
    Run all 4 PDF parsers, pick highest score:
      PyPDF (85%) → PDFPlumber (90%) → LlamaIndex (88%) → Docling (92%) ✓
    Build surface map: what the document CAN and CANNOT answer
    Label each topic: DECISIVE / INFORMATIVE / IGNORE
    │
    ▼
Step 3: Logic/Input → Generator Agent
    Extract: clauses, obligations, key dates, risk flags, governing law
    Uses extended thinking (xhigh) — Claude Sonnet
    │
    ▼
Step 4: Assembly/Execute → Auditor Agent
    Verify every citation exists in the source document
    Check required compliance clauses per document type
    Produce audit score (0–1) and compliance gap report
    │
    ▼
Step 4b: Router Agent
    APPROVED   → confidence ≥ 70% AND audit passed
    REGENERATE → audit failed, regeneration budget remaining (max 2 attempts)
    ESCALATE   → max regenerations hit, low confidence, or unsupported type
    │
    ▼
Step 5: Deploy/Reuse — MCP Server
    extract_legal_document()           → full pipeline
    check_document_quality_tool()      → fast quality gate
    validate_legal_citation_tool()     → citation existence check
    analyze_contract_with_governance() → full + governance rules
    legal_analysis_workflow (prompt)   → workflow prompt for any LLM framework

Quick Start

pip install -e .
cp .env.example .env   # set ANTHROPIC_API_KEY

# Run demo with a synthetic contract (no PDF needed)
legal-agent demo

# Analyze a real PDF
legal-agent analyze contract.pdf

# Quality gate (fast, no LLM)
legal-agent quality contract.pdf

# With governance rules
legal-agent analyze nda.pdf --governance "GDPR compliance, data retention, right to audit"

# HTTP API
legal-agent serve
# → http://localhost:8000/docs

# MCP server (for LangGraph / CrewAI / Google ADK)
pip install -e ".[mcp]"
legal-agent mcp-server

PDF Parsers

Parser	OCR Confidence	Notes
PyPDF	85%	Built-in fallback, always available
PDFPlumber	90%	Highest word count, great for tables
LlamaIndex	88%	pip install -e "[llama]"
Docling (IBM)	92%	Recommended — pip install -e "[docling]"

All four run in parallel. The highest confidence × word_count score wins.

LegalAgentState

from legal_agent import LegalAgentState

# Fail-safe defaults (Dr. Miradi's pattern)
state = LegalAgentState(
    confidence=0.0,    # no trust until measured
    escalation=True,   # escalate to human by default
    validated=False,   # not valid until Auditor passes
)

Accepted Document Types

Type	Accepted	Required Clauses Checked
Contract	✅	governing_law, termination, liability_limitation, payment_terms
Agreement	✅	governing_law, termination, dispute_resolution
NDA	✅	confidentiality_scope, term, return_of_information
Lease	✅	rent_amount, lease_term, security_deposit
Employment	✅	compensation, termination, non_compete
Invoice	❌	Rejected — escalated immediately
Script	❌	Rejected — escalated immediately
Unknown	❌	Rejected — escalated immediately

MCP Tools (reusable across LangGraph / CrewAI / Google ADK)

# Any framework calls the same MCP server
@mcp.tool()
async def extract_legal_document(file_path: str) -> str: ...

@mcp.tool()
def check_document_quality_tool(file_path: str) -> str: ...

@mcp.tool()
def validate_legal_citation_tool(document_text: str, citation: str) -> str: ...

@mcp.tool()
async def analyze_contract_with_governance(file_path: str, governance_rules: str) -> str: ...

Built With

• Anthropic Claude — Generator (extended thinking) + Auditor + surface map
• Docling (IBM) — best-in-class PDF parsing (92% OCR confidence)
• PDFPlumber — table-aware PDF extraction
• FastMCP — MCP server
• FastAPI + uvicorn — HTTP API
• Rich — terminal UI

Based on Dr. Maryam Miradi's framework: www.maryammiradi.com