⚒️ mcp-dataforge

Multi-agent data engineering framework — MCP-native.

Turn natural language into data pipeline actions. Six specialist agents collaborate through the Model Context Protocol (MCP) to build, validate, and monitor your data infrastructure.

---

Quick Start

# Install
pip install mcp-dataforge

# Initialize a project
dataforge init

# Run a task
dataforge run "profile the customers table and check for nulls"

# Start the web dashboard
dataforge web
# → http://localhost:8080

---

Architecture

MCP Client (Claude Code, Cursor, etc.)
        │
        │ MCP Protocol (stdio)
        ▼
┌─────────────────────────────────────┐
│     Orchestrator MCP Server          │
│  route_task · execute_task           │
│  execute_parallel · execute_mixed    │
│  list_agents · get_pipeline_status   │
├─────────────────────────────────────┤
│                                     │
│  ┌──────┐ ┌──────┐ ┌──────┐        │
│  │Pipeline│ │  DQ  │ │Schema│        │
│  └──────┘ └──────┘ └──────┘        │
│  ┌──────┐ ┌──────┐ ┌──────┐        │
│  │Catalog│ │Observ│ │Orch  │        │
│  └──────┘ └──────┘ └──────┘        │
│                                     │
│  Sequential · Parallel · Mixed      │
└─────────────────────────────────────┘

Execution Modes

Mode	Description	Example
Sequential	Agents run one after another, context passes between them	Profile → Detect drift → Generate migration
Parallel	Multiple agents run concurrently, results merged	Scan schema + check health + search catalog
Mixed	Multi-stage: parallel groups followed by sequential steps	[DQ + Schema] in parallel → Catalog

---

Built-in Agents

Agent	Tools	Description
🔧 Pipeline	generate_pipeline, debug_sql, explain_plan	SQL generation, debugging, and optimization
✅ Data Quality	profile_data, detect_anomalies, validate_rules	Data profiling, anomaly detection, rule validation
📐 Schema	detect_drift, generate_migration, lint_schema, lineage	Schema comparison, migration scripts, linting
📚 Catalog	search, describe, impact_analysis, tag	Data discovery, documentation, change impact
🔍 Observability	get_pipeline_health, alert_summary, cost_analysis, suggest_optimizations	Pipeline health, alerts, cost optimization
⚡ Orchestration	create_dag, manage_retry, resolve_deps, backfill, list_dags, pause, unpause, visualize	DAG management, scheduling, dependency resolution

---

CLI Usage

# Project setup
dataforge init                    # Create config.yaml
dataforge agent list              # List configured agents

# Execution
dataforge run "task description"  # Run a one-off task
dataforge start                   # Start orchestrator + agents

# Server modes
dataforge mcp-server              # Run as MCP server (stdio)
dataforge mcp-server --transport sse --port 8080  # SSE mode
dataforge mcp                     # Print MCP config for Claude Code

# Web dashboard
dataforge web                     # Start web UI (http://localhost:8080)
dataforge web --port 9000         # Custom port

Run Complex Pipelines

# Sequential — agents run in order, context flows between them
dataforge run "profile customers table, detect schema drift, and generate migration"

# Multi-agent — single task routed to relevant agents
dataforge run "check data quality and search catalog for PII data"

---

Claude Code Integration

Add to your ~/.claude/settings.json:

{
  "mcpServers": {
    "dataforge": {
      "command": "dataforge",
      "args": ["mcp-server"]
    }
  }
}

Then from Claude Code:

route_task("check null rates in orders table")
→ Returns execution plan with 1 agent (dq)

execute_task("profile customers and fix schema drift")
→ Auto-routes to DQ + Schema agents, runs sequentially, returns results

execute_parallel({"steps": [
  {"agent": "catalog", "task": "search for PII data"},
  {"agent": "observability", "task": "health check"}
]})
→ Both agents run concurrently, results merged

execute_custom_pipeline({"pipeline": [
  {"agent": "dq", "task": "profile orders"},
  {"agent": "schema", "task": "detect drift"}
]})
→ Custom sequential pipeline with context passing

---

Web Dashboard

Start the dashboard to monitor pipelines, agents, and execution history:

dataforge web
# Open http://localhost:8080

Endpoint	Method	Description
/api/agents	GET	List all agents with capabilities
/api/pipelines	GET	List all tracked pipelines
/api/pipelines/{id}	GET	Get pipeline status
/api/execute	POST	Execute a task
/api/pipeline/parallel	POST	Run parallel pipeline
/api/pipeline/custom	POST	Run custom sequential pipeline
/api/pipeline/mixed	POST	Run mixed (parallel + sequential) pipeline

---

Configuration

# config.yaml
version: "1.0"
project: "my-data-platform"

agents:
  pipeline:
    command: "python -m d4.agents.pipeline.server"
    transport: stdio
    capabilities: ["sql", "spark"]
  dq:
    command: "python -m d4.agents.dq.server"
    transport: stdio
    capabilities: ["data_quality", "profiling", "validation"]
  schema:
    command: "python -m d4.agents.schema.server"
    transport: stdio
    capabilities: ["schema", "drift", "migration", "lineage"]
  catalog:
    command: "python -m d4.agents.catalog.server"
    transport: stdio
    capabilities: ["catalog", "discovery", "documentation", "tagging"]
  observability:
    command: "python -m d4.agents.observability.server"
    transport: stdio
    capabilities: ["observability", "monitoring", "alerts", "cost"]
  orchestration:
    command: "python -m d4.agents.orchestration.server"
    transport: stdio
    capabilities: ["orchestration", "dag", "scheduling", "backfill"]

---

Development

# Clone and install
git clone git@github.com:Prometheus-agent/mcp-dataforge.git
cd mcp-dataforge
pip install -e ".[dev]"

# Run tests (153+ tests)
python3 -m pytest

# Run specific test file
python3 -m pytest tests/test_orchestrator.py -v

# Run the MCP server locally
dataforge mcp-server

# Run the web dashboard
dataforge web

Project Structure

src/d4/
├── agents/
│   ├── pipeline/         # SQL pipeline generation
│   ├── dq/               # Data profiling & validation
│   ├── schema/           # Drift detection & migration
│   ├── catalog/          # Data discovery & docs
│   ├── observability/    # Health & cost monitoring
│   └── orchestration/    # DAG management & scheduling
├── config/               # YAML config loader
├── registry/             # Agent registry & discovery
├── orchestrator/         # Core orchestrator + MCP server
├── web/                  # FastAPI web dashboard
├── cli/                  # Click CLI
└── models/               # Pydantic data models
tests/                    # 153+ tests across all modules

---

Building a Plugin

DataForge supports third-party agent plugins:

cp -r templates/d4-plugin d4-plugin-my-agent
cd d4-plugin-my-agent
# Rename <name> to your agent name
pip install -e .

Register in config.yaml:

agents:
  my_agent:
    command: "python -m d4_plugin_my_agent.server"
    transport: stdio
    capabilities: ["my_capability"]

See docs/guides/creating-a-plugin.md for full documentation.

---

Roadmap

Phase 1 — Core Foundation ✅

• 6 specialist agents with 22+ tools
• Orchestrator MCP server (stdio + SSE)
• CLI with init, run, agent, mcp commands
• Sequential, parallel, mixed pipeline execution
• FastAPI web dashboard
• 153+ tests, 100% passing

Phase 2 — Agent Expansion 🚧

• Data Quality agent with DuckDB profiling
• Schema agent with migration generation
• Catalog agent with impact analysis

Phase 3 — Ecosystem 🌐

• Docker deployment
• Plugin API documentation
• Third-party plugin support

---

License

Apache 2.0. See LICENSE.