MCP Integration

Converts the file system tools from the LLM-Powered File System Assistant milestone into a standards-compliant MCP server, and refactors the Agentic Profile Matching LangGraph agent to use those tools through an MCP client instead of a direct Python import.

See state_diagram.md for the full process topology and LangGraph flow.

Part A — filesystem_mcp_server.py

Built on the official MCP Python SDK (FastMCP), so JSON-RPC 2.0 framing, capability negotiation, and request/response dispatch come from the SDK; this file supplies the tool logic, validation, and error semantics.

Tools (the original four, ported from fs_tools.py, plus two new ones):

Resources (discoverable without invoking a tool first):

Error handling: every tool returns {"success": false, "error_code": "...", "error": "..."} on failure (NOT_FOUND, PERMISSION_DENIED, EMPTY_FILE, FILE_TOO_LARGE, UNSUPPORTED_TYPE, MISSING_DEPENDENCY, INVALID_PATH, INTERNAL_ERROR) instead of a bare exception, so a client can branch on failure type.

Configuration (mcp_config.py, overridable via env vars): MCP_MAX_FILE_SIZE_BYTES (100 KB), MCP_CONTENT_CHAR_LIMIT (3000), MCP_SEARCH_MATCH_CAP (10), MCP_BATCH_MAX_FILES (50). Every path a tool touches is resolved and confined to CONFIG.allowed_roots (this project's resumes//data/ plus the sibling milestone projects' resume folders) — new relative to the original fs_tools.py, which trusted its caller; it matters more once any MCP client can reach the server.

Related MCP server: llm-file-operations-agent

Part B — matching_agent.py

Same LangGraph pipeline as the earlier milestone (parse_jd → extract_requirements → search_resumes → rank_candidates → generate_report → human_feedback), but it no longer imports file tools directly. It opens a MultiServerMCPClient (mcp_client.py) that spawns filesystem_mcp_server.py as a stdio subprocess and calls its tools for every file operation — loading the resume corpus via list_files + batch_process, and saving reports via write_file.

Everything else — skill/requirement extraction, scoring, comparison, explanation, interview questions, multi-round screening — is unchanged in-memory logic and needed no MCP involvement.

Bonus: multi-MCP integration

The agent also connects to candidate_db_mcp_server.py, a second MCP server (SQLite-backed) that persists every ranked candidate's verdict after each report. Ask the agent history for <candidate name> to query it back through get_candidate_history. This is deliberately a separate server rather than more tools on the filesystem server — it owns a different kind of resource (database rows, not files), and in a real deployment could run on a different host entirely.

Setup

python -m venv .venv
.venv\Scripts\activate        # Windows
pip install -r requirements.txt

Optional environment variables (same as the earlier milestone):

$env:USE_EXTERNAL_RAG = "1"     # use the ChromaDB index from RAG-Based-Profile-Matching instead of the built-in fallback
$env:USE_GROQ = "1"             # Groq-generated interview questions
$env:GROQ_API_KEY = "your_key_here"

Running

python matching_agent.py

Paste a job description (blank line to submit). Follow-up queries: compare candidates, explain rankings, generate interview questions, history for <name>, add must-haves, next round.

You can also run either MCP server standalone for manual inspection:

python filesystem_mcp_server.py
python candidate_db_mcp_server.py
# or, with the MCP Inspector:
mcp dev filesystem_mcp_server.py

Tests

python test_scenarios.py

9/9 scenarios: the original 7 conversation flows (basic match, natural language skill filter, compare, explain, refine requirements, multi-round, interview questions) plus 2 new ones exercising MCP resource discovery / watch_directory(), and batch_process() combined with the multi-MCP screening-history round trip.

Project structure

MCP Integration/
├── filesystem_mcp_server.py     # Part A — MCP server
├── mcp_config.py                # Server configuration (allowed roots, limits)
├── candidate_db_mcp_server.py   # Bonus — second MCP server (SQLite)
├── mcp_client.py                # MultiServerMCPClient used by the agent
├── matching_agent.py            # Part B — LangGraph agent, MCP client
├── test_scenarios.py            # 9 test scenarios
├── state_diagram.md             # Process topology + LangGraph flow
├── resumes/                     # Sample resumes for standalone server testing
├── data/                        # Generated: latest_match_report.txt, screening_history.db
└── requirements.txt