| name | argo-trading |
| description | Iterate on algorithmic trading strategies in ArgoTradingSwift (`.rxtrading` documents) via the embedded MCP server. Use when the user wants to write, build, backtest, or improve Go-based trading strategies that compile to WASM, load a `.wasm` strategy into the app, run a backtest against a schema/dataset, or analyze backtest results (stats.yaml + per-trade parquet). Triggers include phrases like "write a trading strategy", "backtest this strategy", "beat buy-and-hold", "load_strategy", "run_backtest", and any mention of argo-trading, ArgoTradingSwift, or `.rxtrading` files. |
ArgoTrading Strategy Iteration
Overview
ArgoTradingSwift runs Go strategies compiled to WASM against historical datasets. Iterate on a strategy by editing Go code, building it, loading it through the MCP server, and analyzing the backtest results to beat buy-and-hold.
Prerequisites
Before calling any tool, confirm:
- A
.rxtrading document is open in ArgoTradingSwift (the MCP server always targets the frontmost document window — no document means every tool errors).
- The MCP server is reachable. Default endpoint:
http://127.0.0.1:33321 (probes upward if the port is taken). Stop/change it from Settings → MCP.
get_config returns a selected schema and dataset before run_backtest — otherwise select them first with select_schema / select_data.
Strategy iteration workflow
Follow these steps in order. Do not skip result analysis; without the numbers there is no way to know whether an edit helped.
1. Write or edit the Go strategy
Framework API reference: https://rxtech-lab.github.io/argo-trading/
Keep the strategy name stable across edits. Renaming mid-experiment breaks result comparisons. Pick a name on the first version and reuse it.
2. Build the WASM artifact
From the strategy repo:
make build
This produces a .wasm file. Note the absolute path — load_strategy requires it.
3. Back up a profitable .wasm before rewriting
If the current strategy is green and beats buy-and-hold, copy the .wasm out of the project strategy folder before making a risky edit. Regressions happen; the old binary is the only way back.
4. Load the strategy
Call load_strategy with the absolute .wasm path. It overwrites any previous file with the same name.
5. Select schema and dataset
list_schemas / list_data to discover IDs if unknown.select_schema with a schema UUID.select_data with a dataset filename.- Optionally
read_schema + update_schema to tune backtest_config, live_trading_config, or strategy_config.
- Verify with
get_config.
6. Run the backtest
Call run_backtest. It blocks for up to 5 minutes and returns {status, result_path}.
7. Analyze results (required)
Agents cannot see the chart — do not eyeball it. Always analyze with a Python script that reads stats.yaml and the per-trade parquet.
Use the bundled analyzer:
python3 scripts/analyze_backtest.py <result_path>
It prints stats.yaml, summarizes pnl, and reports whether the strategy beat buy-and-hold. For ad-hoc exploration, load the parquet with DuckDB — see references/analysis.md.
8. Judge against buy-and-hold
The goal is to beat buy-and-hold on the chosen dataset, not just to be green. stats.yaml reports both numbers. If the strategy underperforms buy-and-hold, iterate (return to step 1) — don't ship.
MCP tools
Quick cheat-sheet. Full input/output schemas in references/mcp_tools.md.
| Tool | Purpose |
|---|
load_strategy | Import a .wasm file (absolute path). |
list_schemas / read_schema / update_schema | Discover and edit schema configs. |
list_data | List available datasets. |
select_schema / select_data | Choose what the next backtest runs with. |
get_config | Read currently selected schema + dataset. |
run_backtest | Blocking run; returns result_path. |
Transport is Streamable HTTP JSON-RPC 2.0. For raw HTTP calls (diagnostics only — prefer the MCP tools): POST / with Content-Type: application/json and Accept: application/json to the server endpoint.
Iteration rules (do not violate)
- Keep the strategy name stable across edits.
- Back up profitable
.wasm files before rewriting.
- Always analyze with Python, never by eyeballing the chart.
- The bar is beat buy-and-hold, not just positive PnL.
References
references/mcp_tools.md — full MCP tool input/output schemas.references/analysis.md — DuckDB + yaml patterns for deeper trade-level analysis.scripts/analyze_backtest.py — run this on every result_path before concluding.
