Agentic FinSearch

Agentic FinSearch is a powerful browser extension that combines financial information retrieval with advanced AI capabilities. It provides real-time access to financial data, documents, and insights through an intuitive chat interface.

Contents:

• Overview
  • Key Features
  • Motivation
  • Road Map
• Update Logs
  • Version 0.16.0
  • Version 0.13.3
  • Version 0.13.1
  • Version 0.13.0
  • Version 0.12.0
  • Version 0.11.0
  • Version 0.10.1
• Installation
  • Chrome Web Store (No Local Install)
    • Prerequisites
    • Steps
    • Troubleshooting
  • Quick Install with Docker
    • Prerequisites
    • Steps
    • Updating the Image
  • Prerequisites
  • Install Dependencies with uv
  • Run the Server
  • Frontend Build (optional)
  • Environment Variables
  • Start the Agent on macOS
    • Quick Start (Recommended)
    • Manual Run with uv
    • Frontend Checks
  • Start the Agent on Windows
    • Quick Start (Recommended)
    • Manual Run with uv
    • Frontend Build (optional)
• Using the Agent
  • User Interface
    • Top Bar
    • Main Body
    • Prompt Box
  • Supported Websites
  • Multi-Model Support
    • Available Models
    • Switching Models
  • MCP (Model Context Protocol) Integration
    • MCP Features
  • Deep Research Mode
    • How It Works
  • Custom URL Preferences
    • Setting Preferred URLs
  • Advanced Query Techniques
    • Query Modes
    • Effective Prompting
  • Monitoring Agent Activity
    • Real-time Logs
    • Debug Mode
  • Performance Optimization
    • Smart Context Management
  • Troubleshooting Advanced Features
    • Common Issues
  • Memory and Context System
    • Unified Context Manager (Default)
    • Mem0 Context Manager (Optional)
    • Smart Compression
    • Session Isolation
• XBRL Validation Pipeline
  • Overview
  • The Verification Problem
  • Three Stages
    • Stage 1: Tag the Claim
    • Stage 2: Retrieve the Reference Value
    • Stage 3: Match and Annotate
  • Using Validate
  • Verdict Statuses
  • Current Implementation (Layer 1 MVP)
  • Roadmap
  • Further Reading
• API Reference
  • Connection
    • Base URL
    • Authentication
    • Rate Limiting
    • CORS
  • Endpoints
    • Health Check
    • List Models
    • Chat Completions
    • Available Models
  • Usage Examples
    • Health Check
    • List Models
    • Thinking Mode (MCP Tools)
    • Research Mode (Deep Search)
    • Research Mode with Domain Scoping
    • With URL Context (Page Analysis)
    • Multi-Turn Conversation
    • Normal Mode (No Tools / No Search)
  • Python Benchmarking Quick Start
  • Behavioral Notes
    • Statelessness
    • Response Times
    • Token Usage
    • URL Scraping
    • Error Safety
• MCP Tools Integration
  • Available MCP Servers
  • Architecture
  • How to Enable
  • Configuration
  • Using MCP Tools
  • Troubleshooting
• Project Structure
  • Top-Level Layout
  • Backend Structure
    • Backend Highlights
  • Frontend Structure
    • Frontend Highlights
  • Docker Workflow
  • Manual Backend Workflow
  • Manual Frontend Workflow
• Dev Case Studies
  • Case Study Template
  • Fedora Droplet Storage Exhaustion (01/07/2026)
• Code of Conduct
  • Purpose
  • Scope
  • Core Expectations
  • Unacceptable Behavior
  • Reporting Concerns
  • Confidentiality & Privacy
  • Enforcement
  • Acknowledgment & Updates

Note

This documentation covers version 0.16.0 of Agentic FinSearch.

New in 0.16.0:

• XBRL Validation Pipeline: A three-stage tagging → retrieval → matching pipeline that checks numerical claims in agent responses against SEC XBRL filings and surfaces per-claim verdicts (Verified, Mismatch, Skipped, Not Applicable). See XBRL Validation Pipeline.

• Validate button: Lazy, user-triggered certification. The button appears whenever the most recent response emitted a ratio claim; clicking it runs a deterministic check and inserts inline mismatch marks plus a Ground Truth section in the Sources popup.

• Three Layer 1 axioms: Balance-sheet equality (A = L + E), gross margin, and current ratio — resolved against three pre-loaded SEC filings (AAPL, MSFT, TSLA FY2023). A cloud-hosted SEC XBRL Filing Tree is in development to extend coverage to the full registrant universe.

New in 0.13.3:

• Query Planner & Skills: Heuristic-based skill routing constrains tools per query for faster, more focused responses.

• Hallucination Mitigation: Improved accuracy for financial data retrieval and synthesis.

New in 0.13.1:

• API Now Supported: The OpenAI-compatible API may now be used. For details, please check the API Reference.

New in 0.13.0:

• Deep Research Mode: Multi-step research engine with query decomposition, parallel execution, gap detection, and synthesis.

• OpenAI-Compatible API: RESTful /v1/chat/completions endpoint for programmatic access.

• TradingView MCP: Real-time technical analysis and screener data via Model Context Protocol.

• Gemini Integration: Google Gemini models supported as foundation model providers.

Features

• Multi-Model Support: Choose between OpenAI, Google Gemini, DeepSeek, Anthropic (Claude), and custom fine-tuned models

• MCP Support: Model Context Protocol integration for enhanced agent capabilities

• Browser Extension: Seamless integration with major financial websites

• Real-time Web Scraping: Extract and analyze content from financial websites

Quick Start (Docker)

The simplest way to get started is using Docker:

1. Clone the repository and navigate to the project root.

2. Copy Main/backend/.env.example to Main/backend/.env and add your API keys.

3. Start the application:

   docker compose up --build
   

4. Load the extension in your browser from the Main/frontend/dist directory.

5. Navigate to any supported financial website and start chatting!

Manual Installation

For developers who wish to run or modify the agent outside of Docker, please refer to:

• Prerequisites: Detailed steps for manual backend setup using uv.

• Project Structure: Overview of the codebase and frontend build process.