English | 中文

A local gateway for Claude Code, Codex and other AI clients, on macOS, Windows and Linux. Each client is connected once; after that, upstreams and models change without touching its configuration. Every request is recorded with its cost and route; the API keys in it can be replaced before it leaves the machine, and dangerous tool calls a relay slips into an answer can be cut off before the client runs them.

Sponsors: Want to appear here?

• Connect once, switch freely. Claude Code, Claude Desktop, Codex, opencode, Pi, oh-my-pi, Grok Build, Qwen Code, Hermes Agent, Zed, Aider and DeepSeek Harness are pointed at the gateway in one step, with the change previewed, the original file backed up and a restore always available; Cursor, Continue and Antigravity CLI come with instructions. Switching upstreams then happens in the gateway alone.
• Protection against relays. A relay sees every request in full and can rewrite every answer. Outbound redaction swaps API keys, private keys, JWTs, connection-string passwords, Chinese resident ID numbers and bank card numbers for placeholders before a request leaves, so the relay never holds the real values. When an answer carries a tool call that downloads and runs code, sends out environment variables or credential files, reads private keys or installs a startup item or scheduled job, tool-call inspection cuts the answer off before the client can run it. The content filter deletes instructions hidden in invisible characters before a request leaves and can refuse prompt injection. Each protection starts in Observe, which only records, and is switched over one at a time.
• Upstream check-up. Each upstream is compared with the others serving the same model: answers naming a different model, reported input well above or below theirs and low prompt-cache reads are marked, with sample sizes.
• MCP servers, skills and hooks, scanned. The MCP servers of thirteen clients side by side, with third-party servers marked, and a scan of client configuration, skills, hooks and project instructions for hidden characters, prompt injection, dangerous commands and overly broad permissions.
• Every request traceable. The rule a request matched, each upstream it tried, any conversion between API formats and how its cost was calculated; a finished request can be replayed against another upstream and compared side by side. The whole history can be searched, including the text of requests and answers.
• Routing and failover. Rules by model, tools, images, extended thinking and more. When an upstream fails before the answer begins the next one takes over, and each session stays on one upstream so its prompt cache keeps hitting. Auxiliary requests such as title generation can be answered locally.
• Any upstream, any API format. API keys, Amazon Bedrock, ChatGPT and Z.ai accounts, relays such as OpenRouter and local models, with conversion between the Anthropic, OpenAI and Gemini APIs.
• Costs stated as they are. Estimated amounts are marked and requests without a price are counted separately instead of as zero.
• Remote core. The gateway can also run on a Linux server; the app connects to it over an encrypted control channel.

The gateway, ThinkWatch Core, ships inside the app. The app is not signed by Apple or Microsoft, so the first launch needs one extra step; Install and update covers it, along with how updates arrive. The interface is in English and Simplified Chinese, and the app updates itself (a Homebrew installation updates through Homebrew).

Unzip and run ThinkWatch Lite.exe: no installation and no administrator rights. Configuration, keys and request history stay in the data\ folder next to the program, apart from the installed copy's data, so the two keep separate settings. The folder has to be writable.

Only one of the installed and portable copies runs at a time; opening the other one offers to stop the running one and start it instead. thinkwatch:// links and launch at login always point to the copy that is running.

The portable copy has no system uninstaller. Uninstall it in Settings › Full uninstall first, which restores connected clients and removes its launch at login and other registry entries, then delete the whole folder.

Uninstalling in Settings › Full uninstall restores connected clients, turns off launch at login and can delete the data directory as well. Afterwards, move the app to the Trash on macOS, or delete the AppImage file on Linux.

The installed copy on Windows can also be uninstalled through the system: the uninstaller closes ThinkWatch Lite and does the same restoring and cleanup first. With "Also delete data (configuration, API keys, request history)" ticked it deletes %APPDATA%\ThinkWatch too, except when a client could not be restored, in which case the data directory, with that client's backup, is kept. The uninstaller runs as an administrator: when a standard account uninstalls with an administrator's password, it restores the administrator's clients and deletes the administrator's data instead, so uninstall in the app first in that case.

• Features: every page, in detail
• Install and update
• Connecting to a remote core and server deployment
• Import links, for relays and vendors
• Architecture

pnpm install
bash src-tauri/scripts/fetch-core.sh
pnpm tauri dev

fetch-core.sh downloads the twcore release that Cargo.lock pins into src-tauri/resources/. See CONTRIBUTING.md for the layout of the repository and the checks to run before opening a pull request.

MIT. See LICENSE.