OctoTools is an open-source agentic framework designed for complex reasoning across diverse domains by integrating LLMs with a standardized set of extensible tools. It targets researchers and developers building sophisticated AI agents that require robust tool usage and planning capabilities, offering training-free tool integration and improved performance on complex tasks.

How It Works

OctoTools employs a modular architecture featuring standardized "tool cards" to encapsulate tool functionality, enabling seamless integration of new tools without framework modifications. A hierarchical planner manages both high-level objectives and low-level action refinement. An executor instantiates tool calls, generates executable commands, and stores structured results. The framework also includes a task-specific toolset optimization algorithm to select beneficial subsets of tools for downstream tasks.

Quick Start & Requirements

• Installation: pip install octotoolkit (standard) or pip install -e . (editable after cloning).
• Prerequisites: Python 3.10+, API keys for supported LLM engines (OpenAI, Anthropic, TogetherAI, DeepSeek, Gemini, Grok) and Google Custom Search (API Key, CX). Optional: parallel for benchmark experiments.
• Setup: Requires creating a .env file for API keys.
• Resources: Links to YouTube tutorial and paper preprint are provided.

Highlighted Details

• Achieves 9.3% average accuracy gains over GPT-4o on 16 diverse benchmarks (MathVista, MMLU-Pro, MedQA, GAIA-Text).
• Outperforms AutoGen, GPT-Functions, and LangChain by up to 10.6% with identical toolsets.
• Supports a broad range of LLM engines, including multi-modal capabilities.
• Modular tool card design allows for easy customization and extension.

Maintenance & Community

The project has recent updates (April 2025) including PyPI release and expanded LLM support. A TODO list indicates ongoing development for additional LLM integrations (vLLM, litellm). Community collaboration is encouraged via Slack.

Licensing & Compatibility

The project is licensed under the MIT License, permitting commercial use and integration with closed-source applications.

Limitations & Caveats

Support for vLLM LLM (for custom models) is listed as "in progress." While many LLM engines are supported, specific model availability may vary, and users might need to edit factory.py for custom engine integration.