월스트리트를 겨냥한 '클로드 코드', 랭알파

A vibe investing agent harness LangAlpha is built to help interpret financial markets and support investment decisions. Note Gemini 3 Hackathon — If you're a judge or reviewer for the Gemini 3 Hackathon , please refer to the hackathon/gemini-3 branch for the frozen submission. This main branch contains ongoing development beyond the submission. Getting Started • API Docs • Agent Core • Backend • Web • TUI • Skills • MCP langalpha-demo.mp4 Activated with a skill, the agent dispatches parallel subagents to gather market data, news, and macro context, then presents a morning note with inline interactive visualizations. Why LangAlpha Every AI finance tool today treats investing as one-shot: ask a question, get an answer, move on. But real investing is Bayesian — you start with a thesis, new data arrives daily, and you update your conviction accordingly. It's an iterative process that unfolds over weeks and months: refining theses, revisiting positions, layering new analysis on top of old. No single prompt captures that. From vibe coding to vibe investing Inspired by software engineering: a codebase persists, and every commit builds on what came before. Code agent harnesses like Claude Code and OpenCode succeeded by building agents that embrace this pattern, exploring existing context and building on prior work. LangAlpha brings that same insight: give the agent a persistent workspace, and research naturally compounds. In practice, you create a workspace per research goal ("Q2 rebalance", "data center demand deep dive", "energy sector rotation"). The agent interviews you about your goals and style, produces its first deliverable, and saves everything to the workspace filesystem. Come back tomorrow and your files, threads, and accumulated research are still there. Features Highlights Progressive Tool Discovery — Any MCP tools loaded as summary in context and full documentation dumped into the workspace, allowing the agent to discover and use tools truly on demand. Also supports binding json tools with skills and only expose to agent when skill is activated. Programmatic Tool Calling (PTC) — The agent writes and executes Python to process financial data from mcp servers instead of pouring raw data into the LLM context window, enabling complex multi-step analysis while dramatically reducing token waste. Financial data ecosystem — Multi-tier provider hierarchy with native tools for quick lookups and MCP servers for bulk data processing, charting, and multi-year analysis in sandboxes. Persistent workspaces — Each workspace maps to a dedicated sandbox with structured directories and a persistent memory file ( agent.md ) that compounds research across sessions and threads. Skills for Financial Research — Pre-built workflows for DCF models, initiating coverage reports, earnings analysis, morning notes, document generation, and more — activatable by slash command or auto-detection. Finance Research Workbench — Web UI with inline financial charts, multi-format file viewer, TradingView charting, real-time WebSocket market data, shareable conversations, and subagent monitoring. Multi-provider model layer — Provider-agnostic LLM abstraction and automatic failover on error. Automations — Schedule recurring or one-shot tasks, or set price-triggered automations that fire when a stock or index hits a real-time price condition. Secretary — Flash agent doubles as a secretary: create and manage workspaces, dispatch deep PTC analyses in the background, monitor running tasks, and retrieve results — all through conversational commands with human-in-the-loop approval. Agent swarm — Parallel async subagents with isolated context windows, preloaded toolset/skills, mid-execution steering, checkpoint-based resume, and live progress monitoring in the UI. Live steering — Send follow-up messages while the agent/subagent is working to course-correct, clarify, or redirect without waiting for it to finish. Middleware stack — 24 composable layers handling skill loading, plan mode, multimodal input, auto-summarization, and context management support long-running agent sessions. Security & workspace vault — Encryption at rest via pgcrypto, automatic credential leak detection and redaction, sandboxed execution, and per-workspace secret storage for safe agent access Channel integrations — Use LangAlpha from Slack, Discord with complete feature support. Production-ready infrastructure — SSE-streamed agent activity with Redis-buffered reconnection replay, background execution decoupled from HTTP connections, and PostgreSQL-backed state persistence. What Powers It System Architecture %%{init: {'theme': 'neutral'}}%% flowchart TB Web["Web UI<br/>React 19 · Vite · Tailwind"] -- "REST · SSE" --> API Web -- "WebSocket" --> WSP CLI["CLI / TUI"] -- "REST · SSE" --> API subgraph Server ["FastAPI Backend"] API["API Routers<br/>Threads · Workspaces · Market Data<br/>OAuth · Automations · Skills"] WSP["WebSocket Proxy"] API --> ChatHandler["Chat Handler<br/>LLM Resolution · Credit Check"] ChatHandler --> BTM["Background Task Manager<br/>asyncio.shield · Workflow Lifecycle"] end subgraph PostgreSQL ["PostgreSQL — Dual Pool"] AppPool[("App Data<br/>Users · Workspaces · Threads<br/>Turns · BYOK Keys · Automations")] CheckPool[("LangGraph Checkpointer<br/>Agent State · Checkpoints")] end subgraph Redis ["Redis"] EventBuf[("SSE Event Buffer<br/>150K events · Reconnect Replay")] DataCache[("API Cache<br/>Market Data · SWR")] Steering[("Steering Queue<br/>User Messages Mid-workflow")] end BTM --> AppPool BTM --> CheckPool BTM --> EventBuf BTM --> Steering API --> DataCache BTM -. "Sandbox API" .-> Daytona["Daytona<br/>Cloud Sandboxes"] API -. "REST" .-> FinAPIs["Financial APIs<br/>FMP · SEC EDGAR"] WSP -. "WebSocket" .-> GData["ginlix-data<br/>Polygon.io · Massive"] Loading Multi-Provider Model Layer LangAlpha runs on a provider-agnostic model layer that abstracts across multiple LLM backends. The same middleware stack, tools, and workflows work regardless of which model is driving them. It ships with two modes: PTC mode for deep, multi-step investment research. Strong reasoning drives multi-step analysis where the agent plans its approach, thinks through financial data, and writes code for complex analysis. Long context lets it cross-reference SEC filings and research reports in a single pass. Flash mode for fast conversational responses and workspace orchestration: quick market lookups, chart-and-chat in MarketView, lightweight Q&A, and a secretary that manages workspaces, dispatches deep PTC analyses in the background, and relays results back through natural conversation. Bring your own model — Use your existing AI subscriptions and API keys directly. Connect ChatGPT or Claude subscriptions via OAuth (OpenAI Codex OAuth, Claude Code OAuth), use coding plans from Kimi (Moonshot), GLM (Zhipu), or MiniMax, or supply your own API keys for any supported provider via BYOK. All keys are encrypted at rest via PostgreSQL pgcrypto (see Security ). Model resilience — 3 retries with exponential backoff on the same model, then automatic failover to a configured fallback model. Reasoning effort ( low / medium / high ) is normalized across providers automatically. Programmatic Tool Calling (PTC) and Workspace Architecture Most AI agents interact with data through one-off JSON tool calls which dump the result into the context window directly. Programmatic Tool Calling flips this: instead of passing raw data through the LLM, the agent writes and executes code inside a Daytona cloud sandbox that processes data locally and returns only the final result. This dramatically reduces token waste while enabling analysis that would otherwise exceed context limits. PTC Execution Flow %%{init: {'theme': 'neutral'}}%% flowchart LR LLM["LLM"] -- "1 — Writes Python" --> EC["ExecuteCode Tool"] EC -- "2 — Sends to sandbox" --> Run["Code Runner"] subgraph Sandbox ["Daytona Cloud Sandbox"] Run -- "3 — import tools.*" --> Wrappers["Generated W