fluffypony/dothething

What it solves

dothething is a local AI agent designed to handle complex, multi-step tasks autonomously. It eliminates the need for manual intervention by managing research, data extraction, browser automation, file editing, and code execution until a goal is achieved or a clear failure reason is provided.

How it works

The agent routes various models through OpenRouter, using a primary agent (Claude Fable 5) to plan and execute tasks. It employs a specialized architecture to maintain efficiency and cost:

• Tool Orchestration: It uses "result_mode" to summarize long outputs via Gemini 3.5 Flash and "stacked tool calls" with exec_order to run multiple operations in parallel or sequence in a single round-trip.
• Browser Automation: It uses Notte and Camoufox (a Firefox fork) to bypass bot detection and solve captchas, allowing it to interact with SPAs and complex web forms.
• Hybrid Search: It combines a local SearXNG instance with a browser-based bridge to ensure search engines like Google actually return results, while prioritizing primary-source APIs (like arXiv and PubMed).
• Extensibility: It supports custom skills (via SKILL.md files) and Model Context Protocol (MCP) servers to add new capabilities.
• Resource Management: It implements prompt caching via Anthropic to reduce costs and allows for "Quick mode" for simple, one-shot tasks.

Who it’s for

Users on macOS or Linux who need an autonomous agent capable of performing deep research, managing local files, interacting with the web, and executing shell commands without constant supervision.

Highlights

• Autonomous Web Interaction: Full browser control with fingerprinting avoidance and automated captcha solving.
• Orchestrator Mode: A terminal UI to run and manage multiple agents in parallel.
• Custom Skills & MCP: Ability to inject behavioral instructions or isolated sub-tasks and connect to external MCP servers.
• Integrated Communication: Built-in email capabilities via AgentMail for sending and receiving messages.
• Live Interaction: Supports mid-task input and queued instructions to guide the agent while it works.
• State Persistence: Saves full conversation threads and shell sessions, allowing users to resume interrupted work.