The latest entry in the coding agent wars isn't a CLI tool or a heavy IDE pluginβ€”it's Shelley, a web-based agent that runs on your machine but lives in your browser. Launched with a version number of v0.1346.914604164, Shelley is an open-source, self-modifying agent designed for macOS and Linux. It’s built on a Go backend with SQLite storage and an embedded Vue UI, meaning the entire application ships as a single binary. This architecture allows it to be mobile-friendly and accessible from any device, effectively bringing the coding agent to the compute rather than dragging the compute to the agent.

No Sandbox, Just Raw Access

Shelley takes a hardline stance on security: it comes with no built-in authorization or sandboxing. As the documentation bluntly states, "bring your own." This design choice prioritizes speed and flexibility over the guardrails typical of cloud-hosted agents. Each conversation in Shelley is an isolated agent instance with its own working directory and model configuration, all stored in a single SQLite database. This allows developers to run multiple side-by-side agents, archive old sessions, and search through past interactions without the overhead of complex session management systems.

Multi-Model and Multi-Modal by Design

The agent supports a wide array of LLM providers out of the box, including Anthropic, OpenAI, Gemini, and Fireworks, via environment keys. Users can switch models between turns, leveraging the best tool for each specific task. Beyond text, Shelley is fully multi-modal. It accepts screenshots, charts, and graphs directly into the chat and features a headless browser for navigating and screenshotting web pages. For developers who prefer speaking over typing, Shelley includes voice recording and screen capture capabilities, matching transcripts with video timestamps to give the agent precise context on what the user is pointing at.

Built by Itself, Mostly

Perhaps the most hacker-culture aspect of Shelley is its provenance. The project grew out of an earlier tool called Sketch, and the current codebase is largely written by AI agents themselvesβ€”specifically Shelley, Sketch, Claude Code, and Codex. It integrates deeply with git, offering side-by-side diff viewers, commit tours, and line-level comments. The tool also supports the emerging Model Context Protocol (MCP), allowing users to connect remote servers and drop executable scripts into ~/.config/shelley/hooks to rewrite prompts or react to agent turns dynamically.

Key Takeaways

  • Shelley is a single-binary agent (Go + SQLite + Vue) that runs locally but renders in the browser.
  • It is explicitly unsandboxed, requiring users to manage their own security boundaries.
  • The agent supports multi-modal input (voice, video, screenshots) and multi-model switching.
  • It is open-source under Apache-2.0 and heavily self-authored by AI agents.

The Bottom Line

Shelley is a raw, unpolished gem for developers who want a terminal-free experience without the latency of cloud agents. It’s not for the faint of heart, but for those who want to hack their own hooks and models, it’s the most direct line to an AI pair programmer we’ve seen in a while.