Skip to content

Chrome Faithful v0.4.0 — Exact-profile Chrome control for AI agents

Latest

Choose a tag to compare

@bpc-oss bpc-oss released this 14 Aug 17:06
· 3 commits to main since this release

English

Chrome Faithful provides exact-profile control of real, logged-in Chrome sessions for AI agents through an authenticated localhost bridge, an MV3 extension, and an MCP server.

Highlights

  • Exact-profile Chrome control without copied profiles, a separate debug profile, Edge, or a remote-debugging port.
  • 38 MCP tools, including navigation, interaction, CDP access, network inspection, asset capture, verification helpers, and local visual extraction.
  • First-party DeepSeek Harness bundle targeting DSH 0.1.0-rc.6.
  • Local-first PP-OCRv5 mobile integration. Controlled acceptance detected 7/7 blocks with 99.43% raw and 100% non-whitespace character accuracy.
  • English and Simplified Chinese documentation for the core, DSH bundle, and OCR acceptance report.
  • Transactional Windows installers, authenticated loopback transport, external closed-schema configuration, and DPAPI-protected backups.

Requirements and boundaries

  • Node.js 22.12.0 or newer.
  • Windows is the currently validated host. The Node.js bridge, extension, and MCP server are designed to be platform-neutral.
  • Raw CDP is a trusted-client capability and is not a safe projection for untrusted callers.
  • The optional VLM adapter is disabled by default and has not been quality-approved.
  • DSH model consumption of visual output has not yet been evaluated.
  • npm packages are not part of this GitHub-only release. Install or build from this source release.

Verification: 257/257 tests passed locally. GitHub Actions passed on Ubuntu and Windows, including DSH host isolation, MCPB construction, compatibility checks, and Windows installer transaction tests.

中文

Chrome Faithful 通过经过认证的本地桥接、MV3 扩展和 MCP 服务,让 AI Agent 精确控制用户真实、已登录的 Chrome Profile。

主要能力

  • 无需复制 Profile、额外调试 Profile、Edge 或 remote-debugging 端口,即可精确控制指定 Chrome Profile。
  • 提供 38 个 MCP 工具,覆盖导航、交互、CDP、网络检查、资源捕获、验证辅助与本地视觉提取。
  • 提供面向 DeepSeek Harness 0.1.0-rc.6 的第一方 DSH bundle。
  • 本地优先的 PP-OCRv5 mobile 集成;受控验收中检测 7/7 文本块,原始字符准确率 99.43%,忽略空白后为 100%。
  • 核心项目、DSH bundle 与 OCR 验收报告均提供中英文文档。
  • 提供事务式 Windows 安装器、认证回环传输、源码树外闭环配置与 DPAPI 加密备份。

环境与边界

  • 需要 Node.js 22.12.0 或更高版本。
  • 当前完成实测的平台是 Windows;Node.js 桥接、扩展和 MCP 服务按平台中立方式设计。
  • 原始 CDP 属于受信客户端能力,不应作为面向不可信调用方的安全投影。
  • 可选 VLM 适配器默认关闭,尚未完成质量批准。
  • 尚未验收 DSH 模型对视觉结果的实际消费质量。
  • 本次仅做 GitHub 源码发布,不包含 npm 包;请从本 Release 源码安装或构建。

验证结果:本地 257/257 测试通过;GitHub Actions 的 Ubuntu 与 Windows 作业均通过,包括 DSH 主机隔离、MCPB 构建、兼容性检查和 Windows 安装器事务测试。