veridex-protocol/trading-fabric

This repo is a TypeScript rewrite of an earlier Python project called TradingAgents, where a group of AI agents pretend to be different members of an investment team and argue over what to do with a stock. There are agents that look at the market, social chatter, news, and company fundamentals, then a bull and a bear debate the findings, a trader proposes an action, a risk team challenges it, and a portfolio manager makes the final call. The authors say they kept that team structure but rebuilt the plumbing so the runs are reproducible. Every step of the conversation produces a typed event, the whole run can be saved as a file and replayed later, and the rules about what trades are allowed are written in a separate policy file rather than hidden inside the agent code. Decisions that exceed a daily spend cap or a per-trade position size are escalated to a human approval queue, which is just files in a folder that you accept or reject from the command line. The project ships a command line tool. You can run it on a ticker symbol with a date, replay a saved run, approve or deny a pending decision, inspect the agents memory for a ticker, validate a policy file, or run an evaluation suite. There is also an Ink based terminal UI for watching a run live and a programmatic API for embedding it in other code. By default, trades are written to a paper ledger, but it can be switched to send orders to the Veridex testnet using session keys with budgets. Configuration is done through environment variables and a config object that pick the language model provider (the README lists many options, including OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, GLM, MiniMax, OpenRouter, Ollama, and Azure), set spend caps, choose a data directory, and toggle real execution. There is a Docker Compose setup, docs for architecture, threat model, and policy recipes, and an MIT license. The repo has 0 stars.