fluffypony

dothething

AI🌱 NEW PROJECT BOOST#AI Agent #Automation #LLM #Claude #Python

// summary

Dothething is a local AI agent that autonomously handles complex tasks like research, browser automation, and code execution. It plans its own work, manages tools, and can be extended with custom skills or MCP servers. The system supports persistent sessions, cost tracking, and orchestrator mode for managing multiple parallel agents.

// technical analysis

Dothething (DTT) is a local AI agent designed to autonomously execute complex, multi-step tasks by breaking them down into manageable sub-tasks and utilizing a diverse suite of tools. Its architecture leverages Claude Opus via OpenRouter as the primary decision-maker, while employing specialized models for summarization and oracle-based verification to optimize performance and cost. By integrating browser automation, file manipulation, and custom skill injection, the project solves the problem of manual workflow execution, allowing users to delegate research, coding, and data extraction tasks to an agent that manages its own progress and configuration.

// key highlights

Autonomous task decomposition allows the agent to plan work, track progress, and select appropriate tools without constant user intervention.

Advanced web interaction capabilities utilize Notte and Camoufox to scrape content, solve captchas, and handle complex multi-step browser workflows.

Extensible skill system enables users to define custom behaviors via Markdown files, which can be injected directly into the agent's context or run as isolated sub-tasks.

Orchestrator mode provides a terminal UI to manage multiple parallel agents, including a smart launcher that distributes work across concurrent sessions.

Cost-conscious design features include Anthropic prompt caching, token usage tracking, and the ability to set hard spending limits with stateful checkpointing.

Seamless integration with existing infrastructure is supported through MCP server compatibility and a persistent shell environment for complex build or debugging tasks.

// use cases

Automated web research and browser interaction using Notte and Camoufox

File editing, shell command execution, and code development tasks

Multi-agent orchestration for complex projects with cost and loop limits

// getting started

To begin, clone the repository and execute the dtt.sh script with a prompt describing your task. The system will automatically set up a Python virtual environment and prompt you for necessary API keys, which are securely stored in ~/.dtt/env. You can then explore advanced features by using flags like --orchestrator for parallel tasks or --resume to continue previous work.