mcp-agentic-pipelines

One MCP server. Five AI pipelines. Thirty-one tools. Zero manual setup.

Composable, agentic AI pipelines for Anthropic's Model Context Protocol. Fact-check claims with DSPy, search documents with hybrid RAG, extract intelligence with DeepPipe, and run clinical voice intake — all from a single MCP server. Designed for Claude Desktop, VS Code, Cursor, and any MCP-compatible client.

All 31 Tools

🔍 DeepPipe — Document Intelligence (10 tools)

| Tool | What it does | |---|---| | deeppipe_search | Full-text + vector hybrid search across all documents | | deeppipe_ingest | Ingest raw text/JSON documents for indexing | | deeppipe_ingest_file | Ingest documents directly from file paths | | deeppipe_chat_context | RAG-style Q&A with source citations | | deeppipe_extractive_answer | Extract a precise answer from documents | | deeppipe_list_documents | List all indexed documents with metadata | | deeppipe_get_document | Retrieve a specific document by ID | | deeppipe_get_text | Get the full text content of a document | | deeppipe_remove_document | Remove a document from the index | | deeppipe_stats | View index statistics and document count |

🧠 Piste — Fact-Checking (6 tools)

| Tool | What it does | |---|---| | piste_fact_check | Run a claim through the full 4-stage DSPy pipeline (retrieval → verification → aggregation → verdict) | | piste_list_verdicts | Browse all past fact-check verdicts | | piste_replay | Replay the audit trail of any fact-check run | | piste_get_audit | Get the detailed reasoning chain for a verdict | | piste_get_verdict | Retrieve a specific verdict by claim ID | | piste_submit_feedback | Submit human feedback to improve future checks |

📚 Precis — RAG Pipeline (8 tools)

| Tool | What it does | |---|---| | precis_query | Full RAG query — hybrid search + LLM answer generation | | precis_list_documents | List all documents in the RAG corpus | | precis_debug_stem | Inspect how the stemmer processes a query | | precis_debug_search | See raw hybrid search results (before LLM) | | precis_upload_document | Upload a document to the RAG corpus | | precis_upload_batch | Batch-upload multiple documents at once | | precis_extract_work_order | Extract structured work orders from documents | | precis_list_work_orders | Browse all extracted work orders |

🎙️ Clinical — Voice Intake (5 tools)

| Tool | What it does | |---|---| | clinical_start_session | Start a new voice intake session (returns session ID) | | clinical_process_audio | Send audio → STT transcription → SOAP notes | | clinical_generate_podcast | Generate a clinical podcast from session notes | | clinical_list_sessions | List all voice intake sessions | | clinical_get_session | Get full details of a specific session |

⚙️ Built-in (2 tools)

| Tool | What it does | |---|---| | mcp_health | Server health, tool count, provider status | | mcp_list_providers | List all 9 supported LLM providers |

Compose Pipelines Together

These tools are designed to be chained. Here's a real workflow:

1. deeppipe_ingest           → Load a research paper into the index
2. precis_upload_document    → Add it to the RAG corpus
3. deeppipe_search           → Find relevant passages
4. piste_fact_check          → Verify claims found in those passages
5. precis_query              → Generate a RAG answer from verified context
6. clinical_start_session    → Start voice notes on the findings
7. clinical_generate_podcast → Turn it all into a shareable podcast

In Claude Desktop, this is a conversation:

"Search my documents for claims about climate policy, fact-check each one, then generate a clinical podcast summarizing the verified findings."

Claude orchestrates the tool calls — your server provides the pipelines.

Architecture

graph LR
    subgraph Top[" "]
        direction LR
        CL[Claude Desktop]
        VS[VS Code]
        CU[Cursor]
    end

    subgraph Core["MCP Server Core"]
        direction LR
        ROUTER[Request Router<br/>6 handlers]
        RATE[Rate Limiter]
        GUARD[Dep Guard]
    end

    subgraph DP["DeepPipe<br/>10 tools · TS"]
        D1[search] --> D2[ingest] --> D3[chat]
    end

    subgraph PS["Piste<br/>6 tools · Python"]
        P1[fact_check] --> PY1[DSPy Pipeline]
    end

    subgraph PR["Precis<br/>8 tools · Python"]
        R1[query] --> PY2[RAG Engine]
    end

    subgraph CLN["Clinical<br/>5 tools · TS"]
        C1[session] --> C2[audio] --> C3[podcast]
    end

    subgraph Bottom[" "]
        direction LR
        LLM[9 LLM Providers<br/>DeepSeek · OpenAI · Anthropic · Groq · Ollama · Google · Azure · OpenRouter · Custom]
        EXT[Tavily · Serper · ElevenLabs]
        DEP[uv · 10 pip pkgs · 6 npm pkgs]
    end

    Top --> Core
    Core --> DP & PS & PR & CLN
    DP & PS & PR & CLN --> Bottom

    style Top fill:#f8f9fa,color:#333
    style Core fill:#e8f4fd,color:#333
    style DP fill:#f0f8ff,color:#1a5276,stroke:#aed6f1
    style PS fill:#fef9e7,color:#7d6608,stroke:#f9e79f
    style PR fill:#eafaf1,color:#1e8449,stroke:#a9dfbf
    style CLN fill:#fdedec,color:#922b21,stroke:#f5b7b1
    style Bottom fill:#f8f9fa,color:#333

Quick Start

# 1. Install (npm + Python deps handled automatically)
npx mcp-agentic-pipelines setup

# 2. Start the server
npx mcp-agentic-pipelines

# 3. Or run the test suite
npx mcp-agentic-pipelines test

MCP Client Configuration

Add to your MCP client config:

{
  "mcpServers": {
    "mcp-agentic-pipelines": {
      "command": "npx",
      "args": ["mcp-agentic-pipelines"]
    }
  }
}

Environment Variables

Copy .env.example to .env and set your keys:

| Key | Purpose | |---|---| | DEEPSEEK_API_KEY | Default LLM (DeepSeek) | | OPENAI_API_KEY | Fallback LLM | | GROQ_API_KEY | Clinical voice (STT) | | ELEVENLABS_API_KEY | Clinical voice (TTS) | | TAVILY_API_KEY | Piste web search | | SERPER_API_KEY | Piste web search |

Server starts with any subset — missing keys skip their tools gracefully.

Requirements

• Node.js ≥ 18
• Python 3.11+ (automatically managed via uv — no manual Python install needed on Linux/macOS)
• API keys for the tools you want to use (DeepSeek, Groq, ElevenLabs, Tavily, etc.)

Development

git clone https://github.com/jinan-kordab/mcp-agentic-pipelines.git
cd mcp-agentic-pipelines
npm install
node setup.mjs        # one-time — installs all Python deps
node test.mjs         # runs all 31 tool tests

Architecture (Directory)

packages/
├── core/          Shared types, config, logging, Python bridge
├── server/        MCP entry point, request routing
├── deeppipe/      10 document intelligence tools
├── piste/         6 fact-checking tools (Python bridge → DSPy)
├── precis/        8 RAG tools (Python bridge → hybrid search)
├── clinical/      5 voice intake tools (native TypeScript)
vendors/
├── piste/         DSPy pipeline modules + stdin/stdout bridge
├── precis/        RAG backend modules + stdin/stdout bridge
├── clinical-intake/  Voice pipeline (npm workspace)
setup.mjs          One-time setup (uv, npm, pip)
test.mjs           Complete test suite (31 tools)

License

MIT © Jinan Kordab

Architecture

packages/
├── core/          Shared types, config, logging, Python bridge
├── server/        MCP entry point, request routing
├── deeppipe/      10 document intelligence tools
├── piste/         6 fact-checking tools (Python bridge → DSPy)
├── precis/        8 RAG tools (Python bridge → hybrid search)
├── clinical/      5 voice intake tools (native TypeScript)
vendors/
├── piste/         DSPy pipeline modules + stdin/stdout bridge
├── precis/        RAG backend modules + stdin/stdout bridge
├── clinical-intake/  Voice pipeline (npm workspace)
setup.mjs          One-time setup (uv, npm, pip)
test.mjs           Complete test suite (31 tools)

License

MIT © Jinan Kordab