FinRobot

Overview

FinRobot is an AI Agent platform for financial applications open-sourced by the AI4Finance-Foundation on GitHub, with the official full name "An Open-Source AI Agent Platform for Financial Applications using Large Language Models". It is an agentic AI framework for financial decision intelligence that combines foundation models with financial tools, quantitative models, deterministic computation, and multi-agent workflows to build decision-grade financial applications. Its core principle is: model reasoning, software computation, agent orchestration, system verification. Unlike FinGPT's single-model paradigm, FinRobot supports end-to-end agentic workflows for equity research, quantitative trading, risk analysis, investment banking, private equity, wealth management, and other high-stakes financial decisions.

Project Structure

FinRobot is three projects sharing the same financial domain layer, each serving a different purpose, rather than three competing versions of the same thing.

• V2: The production system, where the deterministic computation engine, provenance guarantees, and agent orchestration are all implemented.
• V1: The equity research web application.
• V0: The AutoGen framework version. Through pip install finrobot what is installed is the V0 package, whose source code has been moved to finrobot_autogen/, but the import name remains unchanged, and all from finrobot... import ... are still available. V0 is retained because it is small enough to read and learn from.

Design Principle: Deterministic Computation + LLM Narration

The design principle running through all generations is: strict separation of deterministic financial computation from LLM-based narration. All financial numbers are generated by pure Python computational operators, not by language models. The LLM is responsible for reasoning, synthesis, interpretation, and report writing; valuation outputs—DCF, DDM, LBO, WACC, comparable company analysis, Monte Carlo simulation—are computed through deterministic code paths with full provenance.

FinRobot Desktop v0.1.0

This is a native desktop equity research cockpit powered by a production-grade multi-agent architecture, guiding analysts from market data and company filings to a single traceable workflow of valuation, debate, synthesis, and investment-committee-style reporting. A design principle visible in the product: the DCF table is labeled "computed by code"—a ten-year forecast generated by pure Python operators; the paragraph beside it is the LLM's interpretation of these numbers across three valuation methods.

Multi-agent architecture: an execution engine driven by a Lead Agent through a pipeline orchestrates specialized research agents, including five role-based sub-agents (data, analysis, modeling, synthesis, reporting), as well as three debate agents (bull case, bear case, judgment). For more details, see finrobot_desktop/README.md. Download and run: for macOS Apple Silicon (M1/M2/M3 or newer), download FinRobot_0.1.0_aarch64.dmg and drag it into Applications. This version does not provide an Intel Mac build. The app has not yet been notarized by Apple, so on first launch macOS may say it is "damaged"; after running a command once in Terminal, it will open normally.

How to Use

V2 has three ways to run the same engine:

• Desktop app: download the release (macOS Apple Silicon).
• Local Web UI: use the same interface in a browser, with no need for the .dmg and Rust toolchain, so it also works on Intel Mac, Linux, and Windows. Running dev.sh will start the FastAPI backend on :8321 and the Vite server proxying the API on :5173; Ctrl+C stops both. Note that it will first free these two ports, so if FinRobot.app is already open, please quit it first; in browser mode the local API is unauthenticated, see finrobot_desktop/README.md.
• CLI: run finrobot_desktop/ in uv sync before use.

For building the Tauri desktop shell, see finrobot_desktop/README.md.

V1 (stock research web app): if deploy.sh cannot run in the environment, there is also a usable two-step CLI pipeline, see finrobot_equity/README.md. Example reports include NVDA, MSFT, TSLA, META, COP.

V0 (AutoGen framework):

1. Create an environment (Python 3.10 or 3.11).
2. Install—from PyPI or from source at the repository root (setup.py maps the finrobot package to finrobot_autogen/finrobot, so installing inside that directory is not feasible).
3. Configure keys—put both files in finrobot_autogen/, where the notebook looks for them. Please copy rather than rename:*_sample files are tracked, and filled-in copies are gitignored.
4. Run the tutorial from finrobot_autogen/tutorials_beginner/ or tutorials_advanced/ .

For the agent library, workflow types, and a complete tutorial index, see finrobot_autogen/README.md.

FinRobot Pro

FinRobot Pro is an AI-driven stock research platform that uses large language models and AI agents to automate professional stock analysis, at https://finrobot.ai/ Features include: automated report generation (generating professional stock research reports on demand), financial analysis (income statement, balance sheet, and cash flow), valuation analysis (P/E, EV/EBITDA multiples, and peer comparison), and risk assessment (comprehensive investment risk assessment).

Research Foundation: Four-Layer Framework

• Financial AI Agent Layer: Includes financial Chain-of-Thought (CoT) prompts to enhance complex analysis and decision-making. The market forecasting agent, document analysis agent, and trading strategy agent use CoT to break down financial problems into logical steps.
• Financial LLM Algorithm Layer: Configures and applies models tuned for specific domains and global market analysis.
• LLMOps and DataOps Layer: A multi-source integration strategy that selects the most suitable LLM for each financial task among a range of advanced models.
• Multi-Source LLM Foundation Model Layer: Supports plug-and-play for general-purpose and specialized LLMs.

Perception-Brain-Action: Perception captures, interprets, and structures multimodal financial data from market quotes, news, and economic indicators; the Brain is the core processing unit, using LLMs to consume perception output and apply financial CoT to generate structured instructions; Action uses tools to execute instructions, turning analysis into trades, portfolio adjustments, reports, or alerts.

Smart Scheduler

Smart Scheduler ensures model diversity and selects the most suitable LLM for each task, including: Director Agent (orchestrates task allocation based on performance metrics and task fit), Agent Registration (manages registration and tracks agent availability), Agent Adaptor (customizes agent capabilities for specific tasks), and Task Manager (stores and manages general-purpose and fine-tuned LLM agents for different financial tasks, and updates them regularly).

Applicable Scenarios and Limitations

Applicable scenarios: stock research, quantitative trading, risk analysis, investment banking, private equity, wealth management, and other high-stakes financial decisions, as well as learning multi-agent financial application development.

Limitations and notes: FinRobot Desktop v0.1.0 only provides a macOS Apple Silicon build, does not provide an Intel Mac build, and has not yet been notarized by Apple, so the first launch may prompt that it is "damaged". In browser mode, the local API is unauthenticated. The official statement clearly declares: the code and documentation here are released under the Apache-2.0 license and should not be regarded as financial advice or live trading recommendations; before any trading or investment decision, one should be cautious and consult a qualified financial professional. The license is Apache 2.0, and for trademark use see TRADEMARK_POLICY.md.