Recoleta
GitHub
Source note

Show HN: Peek-CLI: Let Claude Code See the Browser

Authors
ReactRocks
Source
Hacker News
Published
Collection
Software Intelligence
Coding AgentsBrowser ScreenshotsDeveloper ToolsHuman AI InteractionAgent Safety
Open GitHubMarkdown source

Summary

peek-cli gives coding agents a read-only way to capture screenshots from open browser tabs. It targets agents that need to inspect a running web app without giving them control over the browser.

Problem

  • Coding agents often need visual feedback from localhost apps, UI tests, or browser-based workflows, but many setups do not let them see the browser.
  • Direct browser control can create safety risk because an agent might click, type, inject scripts, or change user data.
  • The tool matters for agent-assisted frontend work because screenshots can help an agent verify UI state and debug visual output.

Approach

  • A Chrome extension captures screenshots from open tabs and streams them through a local WebSocket daemon.
  • The CLI starts the daemon with peeked start, lists visible tabs with peeked list, and saves a tab screenshot with peeked at <url>.
  • Agents use a plugin or skill to request screenshots through the CLI, with setup paths shown for Claude Code and Codex.
  • The security model is read-only: the excerpt claims the agent can request screenshots but cannot access the browser, inject scripts, or perform actions.
  • The user must connect the agent once on each startup, which adds a manual approval step.

Results

  • No benchmark results, user study, latency numbers, or accuracy metrics are provided in the excerpt.
  • The tool claims support for at least 3 named agent clients: Claude Code, Codex, and Copilot.
  • The example flow shows 3 core CLI commands: peeked start, peeked list, and peeked at <url>.
  • The security claim has 1 allowed agent action: capture a screenshot.
  • The project is open source under the MIT license, and the Chrome extension requires manual installation while approval is pending.

Link