sakowicz/actual-ai

What it solves

Actual AI is an automated transaction categorization tool for Actual Budget. It eliminates the manual effort of assigning categories to uncategorized transactions by using Large Language Models (LLMs) to analyze transaction details and suggest the most appropriate category.

How it works

The application connects to an Actual Budget server and identifies transactions that lack a category. It sends the transaction's description, amount, and notes to a configured LLM provider (such as OpenAI, Anthropic, Google, or Ollama). The LLM then classifies the transaction based on the existing category structure.

Key operational features include:

• Web Search Integration: Using the ValueSerp API, the system can search the web for unfamiliar merchants to provide the LLM with better context for categorization.
• Category Expansion: If a transaction doesn't fit existing categories, the LLM can suggest and automatically create new ones.
• Scheduling: It can be run on a cron schedule or at startup to keep budgets up to date.
• Review System: Every guessed transaction is marked in the notes as "guessed," while failed attempts are marked as "not guessed," allowing users to review the changes.

Who it’s for

Users of Actual Budget who want to automate their financial tracking and reduce the time spent on manual transaction classification.

Highlights

• Multi-LLM Support: Compatible with OpenAI, Anthropic, Google Generative AI, Ollama, OpenRouter, and Groq.
• Intelligent Categorization: Uses transaction metadata (payee, amount, date, etc.) and custom prompts to determine categories.
• Web-Enhanced Context: Ability to search for merchant information online to improve accuracy for uncommon businesses.
• Automated Category Creation: Can intelligently suggest and create new categories for transactions that don't fit current groups.
• Safety First: Includes a "dry run" mode to log proposed changes without modifying actual budget data.