Reamer Research now has an MCP server. With it, an AI assistant that speaks the Model Context Protocol, such as Claude Code, Claude Desktop or Cursor, can write a strategy, backtest it with the Reamer Research engine on your own bars, and answer questions from the result: which trades lost the most, what the strategy saw when it entered, how wide the spread of outcomes is. Every run happens on your machine, with your data.
The server, a Claude Code plugin and a Reamer Research skill are in one repository, reamerlabs/reamer-research-mcp, under the MIT licence, and the server is on PyPI as reamer-research-mcp.
What the assistant can do
| Tool | What it does |
|---|---|
engine_status | Finds the installed engine and reports its version, default settings and Evaluation Mode limits. |
api_reference | The strategy API, the order functions and every setting, which the assistant reads before writing code. |
backtest | Runs a strategy file on CSV or .bin bars, with your commission, spread and slippage, and returns the summary metrics. |
closed_trades | Every round trip of a run, sorted by any field: the five worst by net P&L, the longest held, one ticker's. |
monte_carlo | Resamples the run's per-trade returns into loss, ruin and drawdown statistics. |
printed_output | Every value the strategy printed, stamped with its bar's time, so the assistant can explain a decision from the numbers behind it. |
open_trade_diagnostics | Opens the run in visual trade diagnostics: each trade on its own bars, beside every value the strategy printed. |
A strategy stays an ordinary Python file: an on_bar function, a class or an object, the same file rr.backtest runs from a script. The assistant writes it into your project, where you can read it, change it and keep it.
A session
Ask:
> Write an ATR band reversion strategy for Reamer Research, backtest it on SAMPLE_5min.csv with a 0.005 commission and a 0.01 spread, show me the five worst trades and run Monte Carlo.
The assistant checks the engine, reads the API reference, writes atr_bands.py, runs backtest, then calls closed_trades sorted by net_pnl and monte_carlo on the run. Ask why a trade lost and it reads printed_output for the bars around that entry. Ask to see it and open_trade_diagnostics opens the page in your browser. Change a parameter and it runs again; the same data, settings and seed give the same result to the byte, so a difference between two runs comes from the change.
Install
First the engine: download the kit (free, no key) from the download page, then:
tar xzf research-kit-*.tar.gz
pip install research-kit-*/python/reamer_research-*.whl
Then the server. In Claude Code, the plugin adds the server and the skill:
claude plugin marketplace add reamerlabs/reamer-research-mcp
claude plugin install reamer-research@reamerlabs
Or the server alone, in Claude Code:
claude mcp add reamer-research -- uvx reamer-research-mcp
In Claude Desktop, Cursor and other clients, add it to the MCP config:
{
"mcpServers": {
"reamer-research": { "command": "uvx", "args": ["reamer-research-mcp"] }
}
}
The server finds the engine in the Python named by REAMER_PYTHON, else its own Python, else python3 on PATH. If the wheel is in a virtual environment, set REAMER_PYTHON to that environment's python. The MCP server docs cover every tool and setting.
How it runs
Each backtest runs in its own process under the Python that has the engine, so whatever a strategy prints is stored with the run and the connection to the assistant stays clean. That also lets the server run under uvx on its own Python while the engine runs in your environment. Runs are kept in ~/.cache/reamer-research-mcp/runs/, each with the full JSON report, the printed output and the trade diagnostics page, so you can open any of them later.
The skill
The Claude Code plugin also installs a reamer-research skill: how to write a Reamer Research strategy, the order of the workflow above, the engine's rules (an order returned on one bar is eligible from the next, brackets attach to the entry, multi-ticker views gate on valid), and links to the porting guides for Backtrader, NautilusTrader and QuantConnect LEAN. It loads when you ask Claude to backtest a strategy or port one.
Evaluation Mode and a licence
The free download runs every tool in Evaluation Mode: one ticker per backtest, over up to 30 days or 50,000 bars. A run past a limit returns an error naming the limit, and the assistant can narrow the run with start and end. A Reamer Research licence is $20 per seat per month and removes every limit; the same strategy files run unchanged.
Reamer Research is a backtesting and live trading engine for Python and C++, on Linux, macOS, or Windows through WSL2. Its release page is reamerlabs/reamer-research on GitHub.