Skip to content

Agent Quickstart

Connect Claude Code to AlphaForge in about 10 minutes and hand off your first backtest to the agent. We'll walk through installing the CLI, initializing a working directory, and making your first request from Claude Code — in order.

Prerequisite

This page assumes Claude Code is already installed. Other coding agents such as Codex and Cursor work almost identically: the skills that alpha-forge system init scaffolds (below) and the CLI commands are agent-agnostic. If you'd rather connect over an MCP server, see the MCP Reference.

The Trial plan is enough

Everything on this page works end-to-end on the Trial plan — no Whop registration required. Agent integration itself runs fine on Trial. Trial caps data at 2023-12-31, allows up to 50 optimization trials, and disables Pine Script generation (see Getting Started for details).

Step 1 — Install the AlphaForge CLI

If you don't have the CLI yet, install it with a one-liner. You can complete this entire page on the Trial plan.

Run in PowerShell (no admin rights needed).

irm https://alforge-labs.github.io/install.ps1 | iex

After installation, close all open terminals and open a new one before continuing.

curl -sSL https://alforge-labs.github.io/install.sh | bash

After installation, open a new terminal before continuing.

Verify the installation.

alpha-forge --version

AlphaForge, version 0.10.0
Check for updates: alpha-forge self version

For manual installation, custom install paths, or paid-plan login, see Getting Started.

Step 2 — Initialize a Working Directory

Create a working directory for the agent and initialize it with alpha-forge system init. Alongside the usual project layout (forge.yaml, data/, etc.), this scaffolds slash commands and skills for Claude Code and Codex.

mkdir agent-quickstart && cd agent-quickstart
alpha-forge system init

The agent-facing files it scaffolds are:

Path Type Purpose
.claude/commands/explore-strategies.md Claude Code slash command Autonomously explore untried indicator × symbol combos
.claude/commands/analyze-exploration.md Claude Code slash command Aggregate exploration logs and recommend next candidates
.claude/commands/grid-tune.md Claude Code slash command Grid-tune an existing strategy
.claude/commands/tune-live-strategies.md Claude Code slash command Re-tune live strategies
.claude/commands/update-market-data.md Claude Code slash command Batch-update stored historical data
.claude/skills/forge-backtest/SKILL.md Claude Code skill How-to knowledge for running backtests
.claude/skills/forge-analyze/SKILL.md Claude Code skill How-to knowledge for analyzing results
.claude/skills/forge-data/SKILL.md Claude Code skill How-to knowledge for fetching/updating data
.agents/skills/ Codex skills Same skill content as above, for Codex

A one-time EULA prompt appears on first run

The first time you run a primary command, you'll see a EULA summary and a [y/n] consent prompt — once. In non-interactive environments such as agents, set FORGE_ACCEPT_EULA=1 to auto-consent and continue (see "Configuration for Unattended Runs" below).

Step 3 — Your First Request from Claude Code

Launch Claude Code inside the working directory (agent-quickstart/) and ask in natural language. For example:

Backtest SPY with a sample strategy and summarize the results.

Using the scaffolded skills (forge-backtest, etc.), the agent assembles and runs the necessary commands on its own. The flow typically looks like this:

alpha-forge data fetch SPY --period 5y
alpha-forge backtest run SPY --strategy my_rsi_v1 --json

Because it runs with --json, stdout returns pure JSON that the agent reads directly to summarize win rate, max drawdown, Sharpe ratio, and so on (progress and decorations go to stderr, so they never pollute the JSON).

Want to hand off autonomous exploration?

Instead of a single backtest, you can have the agent autonomously explore untried indicator × symbol combinations with the /explore-strategies slash command. It runs backtest → Optuna optimization → walk-forward validation automatically. See the Exploration Workflow for details.

Step 4 (Optional) — Connect the MCP Server

Instead of running the CLI directly, you can expose AlphaForge to the agent over an MCP (Model Context Protocol) server. Register it with Claude Code like this:

claude mcp add --scope user alpha-forge -- uvx alpha-forge-mcp

The difference from running the CLI directly is that each capability is exposed as a typed, safe tool contract (with identifier validation and timeouts built in). alpha-forge-mcp is OSS (Apache-2.0), but it is currently an alpha release, so the tool contract may still change. Cursor and Codex register it with the same command / args. See the MCP Reference for details.

Configuration for Unattended Runs

When you hand off work like overnight autonomous exploration, set environment variables so the agent never hangs on a confirmation prompt.

The three-part set for non-interactive, unattended runs

Setting Role
FORGE_ACCEPT_EULA=1 Auto-consents to the first-run EULA (needed only once)
FORGE_NONINTERACTIVE=1 Removes all confirmation prompts (auto-enabled when CI is truthy or stdout is non-TTY)
--yes / -y Required for destructive operations (delete / purge / prune / clean). Without it, the command exits with code 2 immediately (it does not hang)

To preview a destructive operation first, add --dry-run to show only what would be affected. For the exit-code convention (0 = success / 1 = not found or failure / 2 = argument error) and more, see CLI Conventions.

Next Steps