A trading bot that stages orders like git commits
OpenAlice wraps your broker credentials in an isolated process and makes your AI agent ask permission before it touches real money.

What it does OpenAlice is a self-hosted AI trading agent for equities, crypto, commodities, forex, and macro. It runs two processes on your machine: an “Alice” process that handles research, analysis, and AI-driven decisions, and a separate “UTA” process that actually holds your broker credentials and executes trades. The AI never sees your keys. Every order gets staged, committed with a message, and pushed through a guard pipeline before execution — and you must explicitly approve each trade.
The interesting bit The “Trading-as-Git” workflow is the standout idea. Orders live in a staged state with commit hashes, full history is reviewable, and guards act like ESLint for your portfolio — catching oversized positions, cooldown violations, or off-list symbols before they hit the market. The architecture is explicitly designed to split apart: keep the credential-holding UTA service on a trusted home device while Alice lives on a VPS, like a hardware wallet for trading.
Key highlights
- Multi-broker unification — CCXT, Alpaca, Interactive Brokers, and Longbridge fold into “Unified Trading Accounts” with per-account guard rules and isolated histories
- Research stack built in — TypeScript-native OpenBB engine for cross-asset market data, fundamental equity research, technical indicators, and RSS news archiving
- Workspace-native AI — Per-task git repos with persistent terminal sessions running
claude,codex, or shell, wired via MCP servers rather than protocol shims - Runtime provider switching — Claude (Agent SDK with OAuth) or Vercel AI SDK (Anthropic, OpenAI, Google) without restart
- Self-modification mode — “Evolution mode” escalates permissions to full Bash access, letting the agent rewrite its own code
Caveats
- Explicitly experimental — the README warns against live trading with real funds; many interfaces are incomplete and subject to breaking changes
- Automation’s last mile is unfinished — scheduling (cron, heartbeat) works, but wiring scheduled events into workspace-resident execution is still being rebuilt; heartbeat ticks currently run AI checks without guaranteed user-facing delivery
- Fundamental research depth varies — deepest for equities, other asset classes are expanding
Verdict Worth a look if you want to experiment with AI-driven trading without surrendering broker credentials to a cloud service, and you have the patience for actively maintained alpha software. Avoid if you need turnkey automation or are unwilling to read commit hashes before your orders execute.