AI4Finance-Foundation/FinRobot

FinRobot – an open‑source AI‑agent platform for finance

What it is – FinRobot is a collection of three related code bases that let you build agentic financial‑analysis applications with large language models (LLMs). The project’s philosophy is:

• Models reason – LLMs generate narratives, chain‑of‑thought reasoning, and decide which tool to call.
• Software computes – All numeric finance (DCF, DDM, LBO, Monte‑Carlo, etc.) is performed by pure‑Python functions, guaranteeing deterministic results and full provenance.
• Agents orchestrate – A lead “orchestrator” agent coordinates a pipeline of specialist agents (data collection, analysis, modeling, synthesis, reporting) and a set of debate agents (bull, bear, judge) to produce a traceable research output.
• Systems verify – Audit agents check the numbers, record provenance, and expose confidence scores.

Three versions, one financial‑domain layer

Directory	Version	Intended use
finrobot_desktop/	V2 – production‑grade desktop app (PydanticAI + FastAPI + React/Tauri)	Full‑featured equity‑research workstation; generates 13‑chapter reports, live market data, and deterministic valuation tables.
finrobot_equity/	V1 – self‑hosted web app (OpenAI Agents SDK)	Quick‑start service that turns a ticker into an HTML/PDF research report.
finrobot_autogen/	V0 – educational AutoGen library (pip‑installable finrobot)	Reference implementation and tutorials; not meant for production.

Core design ideas

• Deterministic compute + LLM narration – Numbers come from code; the LLM only writes the story around them. Every output is tagged with its provenance (which operator produced the figure, which data source was used).
• Modular agent stack – The same financial tools can be paired with any emerging agent framework (AutoGen → OpenAI Agents SDK → PydanticAI → future DeepSeek‑Harness).
• Extensive financial tooling – Built‑in data providers (Finnhub, FMP, yfinance, SEC EDGAR, news aggregators, FX), 32 pure‑Python valuation operators, 56 “playbooks” covering equity research, investment‑banking, private‑equity, and wealth‑management workflows.

Getting started (quick overview)

• Desktop (V2) – Download the macOS Apple‑Silicon .dmg or run the source locally:

  cd finrobot_desktop
  uv sync               # install backend deps
  (cd desktop && npm install)  # install frontend once
  ./dev.sh              # launches FastAPI backend + Vite dev server
  

  The UI is reachable at http://localhost:5173 and mirrors the native Tauri app.
• Web app (V1) – After copying config.ini.example and inserting your API keys, run ./finrobot_equity/deploy.sh start to serve the service on http://127.0.0.1:8001.
• Educational library (V0) – Install with pip install finrobot (or pip install -e . from the repo root). Follow the Jupyter notebooks in finrobot_autogen/tutorials_* to see agents generate annual reports, market forecasts, trading strategies, etc.

Example commands (desktop CLI)

finrobot research AAPL          # full 13‑chapter equity research report
finrobot dcf MSFT               # DCF valuation (auto‑switches to DDM if appropriate)
finrobot comps NVDA --peers AMD,INTC
finrobot ic-memo TSLA           # investment‑committee memo
finrobot ask AAPL "How exposed is the gross margin to tariffs?"

All commands produce a structured artifact that includes:

• Data tables (code‑computed, labelled as such)
• Narrative sections (LLM‑generated, referencing the tables)
• Confidence scores and audit notes.

Ecosystem & roadmap

• FinRobot Pro – a hosted SaaS version (https://finrobot.ai/) that offers the same capabilities as a cloud service.
• Future V3 – under development with the DeepSeek‑Harness framework, aiming for more autonomous agents while keeping the verification constraints.
• Research foundation – The repo cites a white‑paper (arXiv 2405.14767) and aligns with recent AI‑agent literature (e.g., Stanford’s Generative Agents).

Who might use it?

• Quant analysts / equity researchers who want an AI‑augmented workflow that still guarantees that the numbers are produced by audited code.
• FinTech developers looking for a ready‑made multi‑agent stack that can be swapped onto different LLM back‑ends.
• Students and researchers interested in learning how to wire LLMs, tool‑calling, and deterministic finance together.

---

Bottom line: FinRobot is a genuine, production‑oriented open‑source platform that combines large language models with a deterministic financial computation engine, packaged as a desktop app, a web service, and an educational library. It stays firmly in the AI‑agent‑for‑finance space.