DSAgent

An AI-powered autonomous agent for data science with persistent Jupyter kernel execution, session management, and conversational interface.

    ____  _____  ___                    __
   / __ \/ ___/ /   | ____ ____  ____  / /_
  / / / /\__ \ / /| |/ __ `/ _ \/ __ \/ __/
 / /_/ /___/ // ___ / /_/ /  __/ / / / /_
/_____//____//_/  |_\__, /\___/_/ /_/\__/
                   /____/

Features

• Conversational Interface: Interactive chat with persistent context and sessions
• Dynamic Planning: Agent creates and follows plans with step tracking
• Persistent Execution: Code runs in a Jupyter kernel with variable persistence across messages
• Session Management: Save and resume conversations with full kernel state
• Multi-Provider LLM: Supports OpenAI, Anthropic, Google, Ollama via LiteLLM
• MCP Tools: Connect to external tools (web search, databases, etc.) via Model Context Protocol
• Human-in-the-Loop: Configurable checkpoints for plan and code approval
• Notebook Generation: Automatically generates clean, runnable Jupyter notebooks
• Agent Skills: Extensible skill system for specialized tasks (EDA, ML, etc.)

Installation

pip install datascience-agent

With optional features:

pip install "datascience-agent[api]"   # FastAPI server support
pip install "datascience-agent[mcp]"   # MCP tools support

For development:

git clone https://github.com/nmlemus/dsagent
cd dsagent
uv sync --all-extras

Docker

Configuration uses the same environment variables as the CLI and server (see Configuration). The container listens on PORT (default 8000).

# Run API server (default: port 8000)
docker run -d -p 8080:8080 \
  -e PORT=8080 \
  -e DSAGENT_DEFAULT_MODEL=gpt-4o \
  -e OPENAI_API_KEY=sk-your-key \
  nmlemus/dsagent:latest

# Run interactive CLI
docker run -it \
  -e OPENAI_API_KEY=sk-your-key \
  -v "$(pwd)/workspace:/workspace" \
  nmlemus/dsagent:latest \
  dsagent chat

# One-shot task
docker run --rm \
  -e OPENAI_API_KEY=sk-your-key \
  -v "$(pwd)/workspace:/workspace" \
  nmlemus/dsagent:latest \
  dsagent run "Analyze data/sales.csv" --data ./data/sales.csv

For Docker deployment details, see docs/DOCKER.md and docs/guide/docker.md.

Quick Start

1. Setup (First Time)

Run the setup wizard to configure your LLM provider:

dsagent init

This will:

• Ask for your LLM provider (OpenAI, Anthropic, Google, local, etc.)
• Store your API key securely in ~/.dsagent/.env
• Automatically select a default model based on provider:
  • OpenAI → gpt-4o
  • Anthropic → claude-sonnet-4-5
  • Google → gemini/gemini-2.5-flash
  • Local → ollama/llama3
• Optionally configure MCP tools (web search, etc.)

To use a different model, set DSAGENT_DEFAULT_MODEL or LLM_MODEL in ~/.dsagent/.env, or use the --model flag:

dsagent --model gpt-4o-mini

2. Start Chatting

dsagent

This starts an interactive session where you can:

• Chat naturally with the agent
• Execute Python code with persistent variables
• Analyze data files
• Generate visualizations
• Resume previous sessions

3. One-Shot Tasks

For batch processing or scripts:

dsagent run "Analyze sales trends" --data ./sales.csv

CLI Commands

Examples

# Interactive chat with specific model
dsagent --model claude-sonnet-4-5

# One-shot analysis
dsagent run "Find patterns in this data" --data ./dataset.csv

# Resume a previous session
dsagent --session abc123

# With MCP tools (web search)
dsagent --mcp-config ~/.dsagent/mcp.yaml

# Human-in-the-loop mode
dsagent --hitl plan

For complete CLI documentation, see docs/CLI.md.

Python API

DSAgent provides two agents for different use cases:

ConversationalAgent (Interactive)

For building chat interfaces and interactive applications:

from dsagent import ConversationalAgent, ConversationalAgentConfig

config = ConversationalAgentConfig(model="gpt-4o")
agent = ConversationalAgent(config)
agent.start()

# Chat with persistent context
response = agent.chat("Load the iris dataset")
print(response.content)

response = agent.chat("Train a classifier on it")
print(response.content)  # Has access to previous variables

agent.shutdown()

PlannerAgent (Batch)

For one-shot tasks and automated pipelines:

from dsagent import PlannerAgent

with PlannerAgent(model="gpt-4o", data="./data.csv") as agent:
    result = agent.run("Analyze this dataset and create visualizations")
    print(result.answer)
    print(f"Notebook: {result.notebook_path}")

For complete API documentation, see docs/PYTHON_API.md.

Supported Models

DSAgent uses LiteLLM to support 100+ LLM providers:

For detailed model setup, see docs/MODELS.md.

MCP Tools

Connect to external tools via the Model Context Protocol:

# Add web search capability
dsagent mcp add brave-search

# Use it in chat
dsagent --mcp-config ~/.dsagent/mcp.yaml

Available templates: brave-search, filesystem, github, memory, fetch, bigquery

For MCP configuration details, see docs/MCP.md.

Session Management

Sessions persist your conversation history and kernel state:

# List sessions
dsagent chat
> /sessions

# Resume a session
dsagent --session <session-id>

# Export session to notebook
> /export myanalysis.ipynb

Output Structure

Each run creates organized output:

workspace/
└── runs/{run_id}/
    ├── data/           # Input data (copied)
    ├── notebooks/      # Generated Jupyter notebooks
    ├── artifacts/      # Charts, models, exports
    └── logs/           # Execution logs

Included Libraries

DSAgent comes with essential data science libraries pre-installed:

Documentation

• CLI Reference - All commands: chat, run, serve, init, mcp, skills
• Configuration - Environment variables and .env
• HTTP API - REST and WebSocket API reference
• Python API - ConversationalAgent and PlannerAgent
• Model Configuration - LLM provider setup
• MCP Tools - External tools integration
• Agent Skills - Extensible skill system
• Docker Guide - Container deployment

License

MIT