cobol-parser-mcp 0.1.1

MCP server that parses COBOL programs — extracts divisions, SQL, CICS commands, and generates AI-powered business logic summaries via any LLM

cobol-parser-mcp

MCP server that parses COBOL mainframe programs and extracts structured data plus AI-powered business logic summaries — works with any LLM provider.

pip install cobol-parser-mcp            # pure parsing, no LLM needed
pip install cobol-parser-mcp[anthropic] # + Claude
pip install cobol-parser-mcp[openai]    # + GPT-4o
pip install cobol-parser-mcp[groq]      # + Groq (fast + cheap)
pip install cobol-parser-mcp[all]       # + all providers

Part of the mainframe modernization pipeline — takes the manifest.json from mainframe-ingest-mcp and produces per-program JSON files ready for code generation.

---

What it extracts

For every COBOL program:

What	Details
IDENTIFICATION DIVISION	Program ID, author, date written
DATA DIVISION	All working storage variables with level numbers and PIC clauses
PROCEDURE DIVISION	Every paragraph with line ranges and PERFORM references
Embedded SQL	Every SELECT/INSERT/UPDATE/DELETE with tables and columns
Embedded CICS	Every SEND/RECEIVE MAP, LINK, XCTL, READ, WRITE
AI summary	Plain-English description of what the program does (any LLM)

---

Claude Desktop config

{
  "mcpServers": {
    "cobol-parser-mcp": {
      "command": "cobol-parser-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "your-key-here"
      }
    }
  }
}

---

Usage from Python

Parse a single file (no API key needed)

from cobol_parser_mcp.tools.parser import parse_cobol_file
import json

result = parse_cobol_file("/path/to/SS6001XX.cob")
print(json.dumps(result, indent=2))

Parse entire codebase with AI summaries

import asyncio
from cobol_parser_mcp.tools.batch import parse_all_programs

# Using Anthropic (Claude)
result = asyncio.run(parse_all_programs(
    manifest_path="./manifest.json",  # from mainframe-ingest-mcp
    output_dir="./parsed",
    ai_summaries=True,
    provider="anthropic",             # or openai, groq, ollama, custom
    api_key="sk-ant-...",             # or set ANTHROPIC_API_KEY env var
    max_ai_programs=20,
))
print(f"Parsed {result['parsed_ok']} of {result['total_programs']} programs")

Using other LLM providers

# OpenAI
result = asyncio.run(parse_all_programs(
    manifest_path="./manifest.json",
    output_dir="./parsed",
    provider="openai",
    api_key="sk-...",
    model="gpt-4o",
))

# Groq — fast and cheap
result = asyncio.run(parse_all_programs(
    manifest_path="./manifest.json",
    output_dir="./parsed",
    provider="groq",
    api_key="gsk_...",
))

# Ollama — fully local, no API key, no internet
result = asyncio.run(parse_all_programs(
    manifest_path="./manifest.json",
    output_dir="./parsed",
    provider="ollama",
    model="llama3",
))

# No AI at all — pure parsing only
result = asyncio.run(parse_all_programs(
    manifest_path="./manifest.json",
    output_dir="./parsed",
    ai_summaries=False,
))

---

Output format

Each program produces a <PROGRAM_NAME>.json file:

{
  "program": "SS6001XX",
  "line_count": 1048,
  "identification": {
    "program_id": "SS6001XX",
    "author": "STATE OF MARYLAND SDAT",
    "date_written": "1995-06-12"
  },
  "data_division": {
    "working_storage": [
      { "level": "1", "name": "WS-ENTITY-RECORD", "type": "GROUP" },
      { "level": "5", "name": "WS-ENTITY-ID",     "type": "PIC X(10)" }
    ],
    "copybooks_expanded": ["BUSENTIT", "SSCC03EQ", "DFHAID"]
  },
  "procedure_division": {
    "paragraph_count": 7,
    "paragraphs": [
      {
        "name": "0000-MAIN",
        "lines": "28-45",
        "calls_paragraphs": ["1000-INIT", "2000-PROCESS", "9999-EXIT"]
      }
    ]
  },
  "sql_statements": [
    {
      "type": "SELECT",
      "tables": ["BUSENTIT"],
      "columns": ["ENTITY_ID", "ENTITY_NAME", "STATUS_CD"],
      "where": "ENTITY_ID = :WS-ENTITY-ID",
      "line": 167
    }
  ],
  "cics_commands": [
    { "command": "RECEIVE MAP", "map": "SS6TMAP", "mapset": "SS6TMAP", "line": 152 },
    { "command": "LINK PROGRAM", "program": "SS6009XX", "line": 334 }
  ],
  "business_logic_summary": {
    "db2_tables_read":    ["BUSENTIT"],
    "db2_tables_written": ["BUSENTIT", "TRNSACTN"],
    "screens_used":       ["SS6TMAP", "SS6XMAP"],
    "programs_called":    ["SS6009XX"],
    "estimated_complexity": "HIGH",
    "ai_summary": {
      "purpose": "Handles online business entity inquiry and status update for CICS terminal users",
      "business_domain": "Business Entity Registration",
      "user_facing": true,
      "key_operations": [
        "Receive user input from SS6TMAP screen",
        "Query BUSENTIT table by entity ID",
        "Update entity status in BUSENTIT",
        "Log transaction to TRNSACTN",
        "Link to SS6009XX for downstream processing"
      ],
      "modernization_notes": "Maps cleanly to GET /api/entity/{id} and PUT /api/entity/{id}/status REST endpoints"
    }
  }
}

A _index.json summary file is also written with stats across all programs.

---

Supported LLM providers

Provider	Default model	API key env var	Notes
anthropic	claude-sonnet-4-20250514	ANTHROPIC_API_KEY	Default
openai	gpt-4o	OPENAI_API_KEY
groq	llama3-70b-8192	GROQ_API_KEY	Fast and cheap
ollama	llama3	none	Fully local, free
custom	gpt-4o	OPENAI_API_KEY	Any OpenAI-compatible endpoint, pass base_url

---

Part of the modernization pipeline

mainframe-ingest-mcp  →  manifest.json
        ↓
cobol-parser-mcp      →  per-program JSON files   ← you are here
        ↓
bms-to-angular-mcp    →  Angular components
db2-schema-mcp        →  PostgreSQL schema

---

License

MIT