Skip to content

alpha-visualizer v1.4.0 — AI strategy development (Agent Develop)

Released: August 3, 2026 / Version: v1.4.0 (this note also covers the v1.4.1 patch) / Distribution: PyPI · GitHub Release

alpha-visualizer is a standalone OSS package that visualizes AlphaForge backtest results in a web browser. v1.4.0 adds Agent Develop — hand strategy development to an AI agent straight from the GUI.

Highlights

1. Agent Develop

Enter a goal, an optional target symbol, and a backend (Claude Code / Codex CLI) on the Develop view (/develop), and your locally installed claude / codex CLI is launched headlessly as an asynchronous job that creates a strategy JSON → validates it with alpha-forge backtest run → links to the new strategy on completion. Observing and cancelling the job uses the same mechanism as the run history.

Security design:

  • No API keys are handled. Authentication and billing stay with each CLI's existing login
  • localhost only. The feature disables itself on non-loopback binds (alpha-vis serve --host 0.0.0.0, etc.)
  • Workspace-scoped. The claude backend scopes its tool allowlist to paths under the workspace and rejects reads/writes outside it (codex uses the OS-level --sandbox workspace-write)

About external communication

This feature launches your own claude / codex CLI as-is. Those CLIs communicate with Anthropic / OpenAI. alpha-visualizer itself never handles, stores, or transmits API keys.

See Develop in Features for usage, the permission model, and turn limits.

2. Turn-limit reporting (v1.4.1)

When the claude backend is cut off at the turn limit (--max-turns), the error now states that fact and the limit value correctly (previously it could be misdiagnosed as an unrelated failure).

How to upgrade

# pip
pip install -U alpha-visualizer

# uv
uv tool upgrade alpha-visualizer

Using Agent Develop requires claude (Claude Code) or codex (Codex CLI) on PATH and authenticated, plus alpha-forge installed. If you don't use it, everything else works as before.

Bug reports and feature requests are welcome on GitHub Issues.