A Self-Hosted AI Trader That Treats Orders Like Git Commits

What it does OpenAlice is a self-hosted TypeScript stack that automates research and trading across equities, crypto, commodities, forex, and macro. It wraps multiple brokers into Unified Trading Accounts, feeds market data and news through a TypeScript-native OpenBB engine, and surfaces everything in a web UI with portfolio dashboards and workspace chat. The agent handles the full lifecycle — research, entry, position monitoring, risk management, and exit — though every order requires your explicit approval before it hits the market.

The interesting bit The architecture splits decision-making from execution like a hardware wallet: the Alice process holds the AI runtime and research tools but never touches broker credentials, while the UTA service owns the keys, order state machine, and a “Trading-as-Git” workflow where you stage, commit, and push orders to execute. That separation is designed so the credential-holding half can eventually run on a home box or phone while the brain lives on a VPS.

Key highlights

• Trading-as-Git workflow with staged orders, commit messages, and an 8-character hash history; pre-execution guard rails enforce position limits, cooldowns, and symbol whitelists per account
• Split-process architecture isolates broker credentials in the UTA service (bound to 127.0.0.1), while the Alice process drives decisions via a thin HTTP SDK
• Multi-provider AI switchable at runtime between Claude (Agent SDK) and Vercel AI SDK (Anthropic, OpenAI, Google) without restarting
• Workspace model pairs a per-task git repo and persistent terminal session (claude, codex, or shell) with MCP-native tool exposure, so the agent works inside a real CLI rather than a chat shim
• Automation scheduling via cron and heartbeat timers is solid, though the execution layer that wires triggers into workspace tasks is still being rebuilt

Caveats

• Explicitly experimental: many interfaces are incomplete, breaking changes are expected, and the authors warn against using it for live trading with real funds
• Fundamental research is deepest for equities and still expanding to other asset classes
• The automation “last mile” — scheduled triggers firing inside a workspace session — is intentionally minimal until the Workspace scheduler lands

Verdict Worth a look if you want to experiment with algorithmic trading without shipping your API keys to a SaaS, but treat it as a sophisticated sandbox rather than a production system. Anyone expecting a polished, fire-and-forget money printer should stay away.