Consulting MCP Server

MCP server that exposes three AI pipelines — RAG Pipeline, Document Intelligence, and Agentic Audit — as 11 composable tools for any MCP-compatible client.

This is the integration layer, not the intelligence layer. The intelligence lives in the pipeline repos. This server makes it consumable through a standard protocol.

Architecture

Tools

RAG Pipeline

Tool	Description
`rag_query`	Single-pass RAG: retrieve + generate grounded answer with citations
`rag_agent_query`	Multi-agent RAG for complex, multi-part questions (slower, more thorough)
`rag_index`	Re-index a corpus directory into the vector store (destructive)

Document Intelligence

Tool	Description
`doc_classify`	Classify a document by type (SOW, Contract, Project Plan, etc.)
`doc_extract`	Full single-doc pipeline: classify + extract structured fields + validate
`doc_assess`	Multi-document assessment with cross-document analysis and narrative
`doc_types`	List available document types and schemas (no API call)

Agentic Audit

Tool	Description
`audit_generate_questions`	Generate interview questions from engagement documents
`audit_process_interview`	Process interview artifacts against a question framework (one app at a time)
`audit_synthesize`	Synthesize all results into an executive summary + Excel deliverable

Utility

Tool	Description
`health`	Server health check: API key, vector store, schemas, pipeline status

Quick Start

Prerequisites

Python 3.12+
Pipeline repos cloned locally (any subset — unavailable pipelines are skipped):
LLM provider config (LLM_PROVIDER, LLM_MODEL, LLM_API_KEY) set in environment or .env

Setup

git clone https://github.com/Brinkv3/consulting-mcp-server.git
cd consulting-mcp-server

python3.12 -m venv .venv
source .venv/bin/activate

# Install server + pipeline dependencies
pip install -r requirements.txt
pip install "llm-adapter[anthropic] @ git+https://github.com/Brinkv3/llm-adapter.git" \
    chromadb sentence-transformers PyMuPDF python-docx \
    python-pptx openpyxl pandas tiktoken

# Configure pipeline paths
cp .env.example .env
# Edit .env with your actual paths and API key

Connect to Claude Desktop

Copy the config into your Claude Desktop settings (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "consulting-mcp-server": {
      "command": "/path/to/consulting-mcp-server/.venv/bin/python",
      "args": ["src/server.py"],
      "cwd": "/path/to/consulting-mcp-server",
      "env": {
        "RAG_PIPELINE_PATH": "/path/to/rag-pipeline",
        "DOC_INTEL_PATH": "/path/to/doc-intelligence",
        "AUDIT_PATH": "/path/to/agentic-audit",
        "LLM_PROVIDER": "anthropic",
        "LLM_MODEL": "claude-sonnet-4-6",
        "LLM_API_KEY": "your-key-here"
      }
    }
  }
}

See config/claude_desktop_config.json for a complete example.

Connect to Claude Code

claude mcp add consulting-mcp-server \
  -e RAG_PIPELINE_PATH=/path/to/rag-pipeline \
  -e DOC_INTEL_PATH=/path/to/doc-intelligence \
  -e AUDIT_PATH=/path/to/agentic-audit \
  -- /path/to/consulting-mcp-server/.venv/bin/python src/server.py

Verify

Once connected, ask Claude to run health — it reports the status of each component:

Server: running
RAG pipeline: available
Doc intelligence: available
Agentic audit: available
LLM provider: anthropic
LLM API key: set
Vector store: found
Schemas: found (6 types)

Architecture

MCP Client (Claude Desktop / Claude Code / any MCP client)
      │ (MCP protocol over stdio)
      ▼
consulting-mcp-server
  ├── server.py         → MCP server entry point, tool registration
  ├── rag_tools.py      → Tool handlers wrapping RAG pipeline
  ├── doc_tools.py      → Tool handlers wrapping doc intelligence
  ├── audit_tools.py    → Tool handlers wrapping agentic audit
  └── utils.py          → Config, path validation, pipeline imports
      │                    │                    │
      ▼                    ▼                    ▼
  RAG Pipeline        Doc Intelligence     Agentic Audit
  (path-based import) (path-based import)  (path-based import)

All three pipelines use src/ as their package name. The server imports them sequentially, flushing sys.modules between imports to avoid namespace collisions. Each pipeline is optional — if a path isn't configured, its tools report "unavailable" and the rest of the server works normally.

Tests

source .venv/bin/activate
pytest tests/ -v