OpenAlice is drawing attention across crypto and quant trading circles with a simple pitch: run an AI trading desk on a laptop, keep the code open, and keep the data local. Created by @0xcherry, the project has recorded more than 3,100 GitHub stars, 425 forks, and 463 commits. Its public positioning is blunt—research, execution, and risk controls in one locally operated system.
Research, analysis, and order staging in one stack
Unlike many trading bots that rely on fixed rules and static parameters, OpenAlice is built around continuous reasoning. The project includes a modified OpenBB-based engine called TypeBB, which can work with equities, crypto, forex, options, futures, ETFs, macroeconomic indicators, and BLS labor data. It also supports spreadsheet-style formulas for indicators such as moving averages, RSI, and Bollinger Bands. On the information side, the system can monitor selected RSS feeds, archive incoming items, analyze them, and judge whether a development is bullish or bearish for an existing position.
Execution is handled with human approval built in. OpenAlice supports more than 100 crypto exchanges, including Binance, OKX, and Bybit, and it also covers US stocks, equity options, futures, and bonds. Orders are first placed into a staged state, then reviewed by the user, then committed and pushed to a broker. It does not send trades straight to market on its own.
Local-first design and account-level guardrails
Risk control is a central part of the project’s structure. Each account can be assigned its own Guard Pipeline, where users define limits such as maximum position size per trade, cooldown periods, and allowed instruments. If a proposed trade fails those checks, it cannot be sent out.
OpenAlice also uses a file-driven architecture. Its state is stored in a local data directory with Markdown files, JSON configuration files, and conversation logs that users can read directly. The project avoids database dependence and cloud sync, and it can connect to locally hosted large language models so model-level data can remain on the same machine as well.
The system is designed for persistence, not one-off prompts. According to the project description, Alice keeps a persistent cognitive state, retains conclusions from earlier sessions, and tracks an internal emotional state intended to shape decision style over time.
Use cases range from crypto monitoring to natural-language trade instructions
The examples provided by the project cover several workflows. One is crypto swing monitoring, such as tracking 4-hour BTC and ETH charts alongside selected news feeds and then generating analysis and position suggestions when key headlines or breakouts appear. Another is earnings-season support for US equities, where the system can collect analyst expectations, review prior reports, and flag dates for companies already held in a portfolio.
Users can also issue trade requests in plain language through the web interface—for example, reducing an NVDA position by 30% over three days at average daily pricing. In another setup, Alice checks portfolio risk every hour and alerts the user through iMessage or Telegram if a position drawdown exceeds 5%.
Because the agent can call native operating-system functions, it can search through a browser, send iMessage notifications, and connect to local network speakers for alerts. The project presents this as a native system agent rather than a narrow sandbox tool.
Open-source momentum comes with a clear risk warning
To get started, users need Node.js 22 or higher, pnpm 10 or higher, and a logged-in Claude Code CLI environment. The source article says users familiar with command-line setup may get it running in about two hours, while others may spend much longer on configuration.
OpenAlice was open-sourced in early 2026. The creator said maintaining a commercial product meant spending time on pricing plans, production systems, and permissions instead of improving Alice itself. The article notes three possible business directions under consideration: data subscriptions, hosted services, and transaction fee sharing.
The project also makes its limits clear. It is described as experimental software, and the quality of its decisions depends on the underlying large language model. Hallucinations, misleading news, and weakening signal quality in technical indicators remain real risks. No backtest data is cited in the source material, and the official guidance does not recommend using real funds unless the user fully understands and accepts the risks.