// the find
ginlix-ai/LangAlpha
Claude Code for Investing
LangAlpha is a financial research agent harness built on LangGraph, giving investors persistent workspaces where an AI agent can compound analysis across sessions rather than answering one-off questions. It targets quant-adjacent developers and finance professionals who want something closer to a research environment than a chatbot. The core bet is that investment research is iterative and stateful, and that agentic AI fits that shape better than one-shot prompting.
Programmatic Tool Calling (PTC) is the genuinely interesting idea here: instead of pouring raw financial data into the context window, the agent writes Python code that runs in a sandbox and returns only the processed result. This is a real solution to a real problem — financial time series and fundamentals data will blow your context budget fast. The persistent workspace design with agent.md injected into every call is well thought out; research does compound across sessions in a way that chat history alone does not support. Infrastructure is not a toy: SSE event buffering with 150k-event replay, LangGraph checkpointing in PostgreSQL, and background tasks decoupled from HTTP connections means the agent keeps working if your browser tab closes. The three-tier data provider fallback (ginlix-data → FMP → Yahoo Finance) lets you start free and upgrade incrementally, which is pragmatic.
The differentiating feature — PTC with cloud sandboxes — requires Daytona, and without it you drop to a Docker fallback the README describes as 'downgraded security and isolation.' Price-triggered automations are beta-only on their hosted service. The architecture has real vendor tie-in that is easy to miss reading the feature list. LangChain and LangGraph are load-bearing here, and both libraries are notorious for abstraction leakage and opaque failure modes; the additional middleware stack on top means debugging a stuck long-running agent session is going to be painful. The full setup is genuinely complex — workspace vault, MCP servers per workspace, three data tiers, multiple optional keys — and the free-tier experience (Yahoo Finance data, no intraday quotes, no real-time feed) is materially worse than what the screenshots show, which is undersold. There is no evaluation or backtesting layer: the agent outputs DCF models and initiating coverage reports, but nothing in the repo lets you assess whether those outputs correlate with actual market outcomes, which is the only thing that would matter if you were using this seriously.