Local Claude Bridge 0.4.2

Local Claude Bridge 0.4.2 for Windows x64.

• Discovers both the main Claude model menu and its More models submenu, including Sonnet 5.5 and Haiku 5.5 when selectable by the signed-in account. Upgrade-only entries are excluded.
• Passes low, medium, high, xhigh, and max effort to the web endpoint with thinking_mode extended. OpenAI reasoning_effort / reasoning.effort and Messages-shaped thinking / output_config.effort override the saved default. Effective controls appear in X-Local-Claude-Thinking.
• Settings tabs: Account, API, Models, Keys, App, and Logs. Click the top-right endpoint URL to copy it.
• Themed right-click menus include model, thinking, dark mode, browser visibility, startup, API options, update checks, and settings shortcuts.
• Automatic session checks, first-workspace selection, optional endpoint resuming, connection/API indicators, and a hidden-browser dashboard.
• Proper shared-profile sign-in popups preserve window.opener instead of replacing the main Claude page.
• Custom icon, single Windows installer EXE, portable ZIP, and verified in-app updates.

Validation: 56 API assertions, 28 updater assertions, JavaScript stream/transport/model-discovery suites, and a real WebView2 synthetic popup regression passed locally. Live model discovery returned Sonnet 5.5 and Haiku 5.5 along with other available models. A live Sonnet 5.5 request with High thinking returned HTTP 200 and OK. Earlier installation tests preserved key hashes, port, model, and authentication mode. Full visual and small-screen checks remain incomplete.

Recommended download: LocalClaudeBridge-Setup-0.4.2-x64.exe. Installs for the current Windows user, creates a Start-menu shortcut, and opens the app. .NET is bundled; Microsoft Edge WebView2 Evergreen Runtime is required. The custom installer does not register an Add/Remove Programs uninstaller. The ZIP can be extracted and run as a portable app; keep its updater helper with the application.

The adapter uses undocumented website interfaces. It remains experimental, uses buffered streaming and prompt-based tool translation, and does not provide exact token budgets, returned thinking blocks, unlimited usage, or quota bypass. A successful thinking-enabled request confirms upstream acceptance, not an exact reasoning-token budget.

Public MIT source. Inspiration and references: EvanZhouDev/openai-oauth, mattbnz/LLMux, and the OmniRoute Claude web adapter. See README.md and LICENSE.

Release-package verification: the installer installed and launched version 0.4.2 from commit 316211e. The settings file SHA-256 remained unchanged, the browser profile was retained, the saved Claude session verified, all five selectable models were discovered, and the local endpoint resumed successfully. Uploaded ZIP, installer, and checksum digests and sizes matched local files.