Most browser automation tools start with a human-first browser and bolt on scripting. Vessel flips that assumption: the AI agent is the primary operator, and the human watches from a supervisor sidebar.
Built by solo developer Tyler Williams at Quanta Intellect in Portland, Oregon, Vessel came out of the Nous Research Hermes Agent Hackathon and is now available as an MIT-licensed open-source project on GitHub.
How It Actually Works
Vessel runs on Chromium but strips away the assumption that someone is sitting there clicking. Instead, it exposes 40+ specialized tools to AI agents through MCP (Model Context Protocol, the emerging standard for connecting AI models to external tools). An agent can call structured commands like navigate, click, type, extract, and save_session without needing to parse raw HTML or guess at pixel coordinates.
The browser extracts semantic context from pages - visibility states, disabled elements, dropdown options - and feeds that to the agent. This is a real improvement over screenshot-based approaches where agents try to interpret pixels like a human would.
A demo on the project site shows Vessel autonomously browsing Powell's bookstore, finding five books, and adding them to a cart with zero human clicks.
The Supervisor Model
The human oversight piece is well thought out. A sidebar shows every action the agent takes in real-time, and you can set approval policies per action type: auto-approve routine navigation, require approval before form submissions, deny certain actions entirely. You can also pause and resume the agent mid-task.
Sessions persist across restarts with named session states and checkpoint snapshots, so an agent can pick up where it left off after a crash or reboot.
Who This Is For
Vessel supports multiple LLM providers out of the box - Anthropic, OpenAI, Ollama, Mistral, xAI, Google Gemini, and OpenRouter - with a bring-your-own-keys setup. It currently runs on Linux only, available as an AppImage, an npm package (@quanta-intellect/vessel-browser), or built from source.
This sits in a different lane from tools like Playwright or Puppeteer, which automate browsers for testing. Vessel is specifically designed for open-ended agent tasks where the AI needs to reason about what to do next, not just follow a script. It also differs from cloud-based agent browsers by running locally, which matters for anyone handling sensitive data or internal tools.
As a solo-developer hackathon project, expect rough edges. But the architecture - MCP-native, semantic context extraction, granular human oversight - reflects where agent-controlled browsing is heading. The Linux-only limitation will narrow the audience for now, though npm and AppImage installs keep the barrier low for those on the right OS.