Freqtrade Dev MCP

MCP server that exposes Freqtrade strategy development tools (data download, backtesting, hyperopt, result analysis) to LLM clients like Claude Desktop and Claude Code. It ships with a LangGraph agent that uses the same tools to generate, optimize and evaluate strategies on its own.

Independent project, not affiliated with the Freqtrade team. For research and education. Backtest results do not predict live performance. Never trade real funds with a strategy you have not reviewed yourself.

What's inside

MCP tools

Full parameter reference: docs/tools-reference.md.

Requirements

• Python 3.11+ (required by freqtrade>=2025.7)
• A Freqtrade installation with a user_data/ directory
• freqtrade on PATH for backtest/hyperopt runs
• For the agent: an API key for one of the supported LLM providers

Installation

git clone https://github.com/dasein108/freqtrade_dev_mcp.git
cd freqtrade_dev_mcp

# MCP server only
uv pip install -e .

# MCP server + strategy agent
uv pip install -e ".[agent]"

# Development tools
uv pip install -e ".[agent,dev]"

pip install -r requirements.txt also works and installs everything, agent included.

Configuration

The server reads ~/.config/freqtrade-mcp/config.json if present, then applies environment overrides:

File-writing tools (create_userdir, create_config, create_strategy*) only accept paths inside FREQTRADE_MCP_PATH; relative paths resolve against it. New strategies go to user_data/strategies/ by default, where backtest and hyperopt look for them.

Example config file:

{
  "freqtrade_path": "/path/to/freqtrade",
  "default_exchange": "binance",
  "default_stake_amount": 100.0,
  "default_epochs": 100
}

Connecting an MCP client

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "freqtrade": {
      "command": "/absolute/path/to/python",
      "args": ["/absolute/path/to/freqtrade_dev_mcp/run_server.py"],
      "env": {
        "FREQTRADE_MCP_PATH": "/absolute/path/to/freqtrade"
      }
    }
  }
}

Restart Claude Desktop afterwards.

Claude Code

claude mcp add freqtrade -e FREQTRADE_MCP_PATH=/absolute/path/to/freqtrade \
  -- /absolute/path/to/python /absolute/path/to/freqtrade_dev_mcp/run_server.py

More examples in examples/.

Typical workflow

Ask your MCP client in plain language; it maps the request to tool calls like these:

download_candles(pairs=["BTC/USDT", "ETH/USDT"], timeframes=["1h"], date_range="last 6 months")
create_strategy_wireframe(strategy_name="EmaTrend", style="guided",
                          description="EMA crossover trend follower")
backtest_strategy(strategy_name="EmaTrend", pairs=["BTC/USDT", "ETH/USDT"], timerange="last 3 months")
hyperopt_strategy(strategy_name="EmaTrend", pairs=["BTC/USDT", "ETH/USDT"],
                  timerange="last 6 months", epochs=200, spaces="buy sell")
extract_hyperopt_data(hyperopt_path="/path/to/EmaTrend.fthypt", output_format="summary")
search_results(result_type="backtest", min_profit=5, max_drawdown=15, sort_by="profit")

To limit overfitting, validate on a time range that hyperopt never saw (walk-forward).

Strategy agent

The agent runs the full loop without a human in the middle:

generate idea → write strategy code → download data → hyperopt → backtest → analyze
      ↑                                                                    │
      └──────────────── rewrite if below profit threshold ─────────────────┘

Set up .env (see .env.example):

LLM_MODEL=openai/gpt-4o-mini       # provider/model: openai, anthropic, deepseek, groq, together, ...
LLM_API_KEY=...
# optional
LLM_TEMPERATURE=0.3
MAX_ITERATIONS=3
HYPEROPT_EPOCHS=100
MIN_PROFIT_THRESHOLD=5.0

Run it:

python run_strategy_agent.py \
  --symbols BTC/USDT:USDT ETH/USDT:USDT \
  --timeframes 1h \
  --max-iterations 3 \
  --hyperopt-epochs 100 \
  --min-profit 5.0

Details: docs/strategy-agent.md and docs/agent-architecture.md.

Project structure

freqtrade_dev_mcp/
├── src/                     # MCP server
│   ├── commands/            # One module per tool (BaseCommand subclasses)
│   ├── models/              # Pydantic response models shared with the agent
│   ├── utils/               # Date parsing, CoinGecko, logging helpers
│   ├── config.py            # Config file + env loading
│   └── server.py            # Tool registration and dispatch
├── strategy_agent/          # LangGraph agent
│   ├── nodes/               # data_fetcher, strategy_generator, hyperopt_runner, result_analyzer
│   ├── prompts/             # LLM prompt templates
│   ├── agent.py             # Graph definition
│   ├── mcp_client.py        # stdio MCP client
│   └── llm_client.py        # Multi-provider LLM factory (instructor)
├── tests/                   # unit/ and integration/
├── examples/                # Client configs and scripts
├── scripts/                 # Maintenance scripts
├── docs/                    # Documentation (dev-notes/ = historical fix logs)
├── run_server.py            # MCP server entry point
└── run_strategy_agent.py    # Agent CLI

Development

make install-dev   # install deps
make test          # pytest
make lint          # ruff
make format        # black + ruff --fix
make type-check    # mypy

Server logs go to stderr and logs/mcp_server_*.log (falls back to $TMPDIR/freqtrade_mcp_logs/ if logs/ is not writable). Agent logs go to logs/strategy_*.log (plain text and JSON). python scripts/check_imports.py does a quick import sanity check.

Documentation

• Tools reference: every tool, parameter and response
• Developer guide: architecture, data models, adding tools, testing
• Strategy agent: agent usage, configuration, workflow
• Agent architecture: original LangGraph design notes
• Tests: how the test suite is organized
• Changelog

Project status

Alpha. Tool interfaces and response schemas may still change. See open issues for known problems.

Contributing

1. Branch from dev.
2. Add tests for new behavior (pytest).
3. Run make lint type-check test.
4. Open a pull request against dev.

Coding conventions live in CLAUDE.md.

Support

• Bugs, questions and feature requests: GitHub Issues

License

MIT. See LICENSE.