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. For macOS, Windows and Linux, under the MIT License.
Sponsors: 
Become a sponsor
Set up in one step, or by following the instructions
01Routing
Rules send requests to different upstreams by model, tools, images, extended thinking and more, and can set the reasoning effort. When an upstream fails before the answer begins, the next one takes over, a brief failure is retried after a short wait, and each session stays on one upstream so its prompt cache keeps hitting. Auxiliary requests such as title generation and warm-ups can be answered locally without using any quota.
02Protection
API keys, private keys, connection-string passwords, ID numbers and bank card numbers can be replaced with placeholders before a request leaves, so a relay never holds the real values, and an answer whose tool call downloads and runs code or sends out credentials can be cut off before the client runs it. Hidden characters that smuggle in instructions can be deleted, prompt injection refused, and chosen words deleted or replaced in requests and answers. The three protections start in Monitor, which records and notifies without changing anything; each is switched on separately, and every rule has its own action.
03Tracing
A request shows the rule it matched, each upstream it tried, any conversion between API formats and how its cost was calculated, and its content can be read live while the answer streams. A finished request can be replayed against another upstream and the two answers compared side by side, and the whole history can be searched, including the text of requests and answers.
04Cost
Tokens, cost, cache hits, requests, time to first token and traffic on one screen, each against the previous period and broken down by model, upstream or key. Estimated amounts are marked, and requests without a price are counted separately instead of as zero; prices follow LiteLLM's public list, refreshed daily, or a custom price sheet.
The macOS menu bar shows today's tokens and cost. Its menu gives the gateway's state, each quota with its reset time and the requests in progress, and copies the gateway address or default key without opening the window. Windows and Linux have the same menu in the tray.
12
set up in one step · 3 more by instructions
Claude Code, Codex, opencode and nine other clients are pointed at the gateway in one step, with the change previewed, the original file backed up and a restore always available; on Windows, Claude Code and Codex inside WSL as well. From then on, switching upstreams happens in the gateway, with no client to reconfigure or restart.
API keys, Amazon Bedrock, ChatGPT and Z.ai accounts, relays such as OpenRouter and local Ollama models all serve as upstreams, with the balances, quotas and reset times their services report shown. Requests are converted between the Anthropic, OpenAI and Gemini APIs, so Codex can also use models that only speak Chat Completions.
The MCP servers of thirteen clients appear side by side, with third-party remote servers and inconsistent configurations marked, and can be copied or removed between clients. Client configuration, skills, hooks and project instructions are scanned for hidden characters, prompt injection, dangerous commands and overly broad permissions, and a new finding raises a notification.
Connecting a client gives it a key of its own, so traffic and cost are counted per client. Each key has its own route, visible models, concurrency limit and usage limits on requests, tokens or cost, and a rotated key is written into its client's configuration.
A dry run shows which rule a request would match, why the rules before it did not, and which upstreams would be tried in turn. Nothing is sent and nothing is charged.
A system notification reports a gateway that stopped forwarding, a lost remote connection, a quota that ran out, a credential that stopped working, an unreachable proxy, configuration that did not take effect, a dangerous tool call that was cut off, and suspicious content in client configuration.
Short JavaScript plugins adapt requests and answers to a particular setup; the app includes two, one for answers in a chosen language and one for converting file paths between WSL and Windows. Each plugin runs in a sandbox without network or file access, reads and changes only the parts its permissions declare, and receives placeholders instead of keys; every change is recorded on the request and checked by the same protections as anything a client sends.
Remote core
The gateway also runs on a Linux server as a systemd service, serving clients across the network. The app connects to it and shows the server's traffic, cost and configuration in the same pages, and the clients on the computer can be pointed at the server's gateway in one step.
The control connection is encrypted and authenticated by a Noise handshake, with no certificates involved. The app connects to one core at a time, and the local data stays as it was while it is connected elsewhere.
Architecture
Lite holds no routing, forwarding or accounting logic. All of it is in ThinkWatch Core, which runs beside the app or on a Linux server.
Desktop app
Gateway
Upstreams
The control channel is a Unix socket on macOS and Linux, a loopback port on Windows and a TCP port to a core on a server; every connection is encrypted and authenticated by a Noise handshake.
Install
The gateway ships inside the app; nothing else needs to be installed.
Requires macOS 12 or later on Apple silicon.
Homebrew installs the app and keeps it up to date.
$ brew install --cask thinkwatchproject/tap/thinkwatch-lite Download, open and drag the app into Applications. The app then updates itself.
The app is not notarized by Apple. When it is installed from the disk image and macOS blocks the first launch, choose Open Anyway in System Settings › Privacy & Security. The Homebrew install needs no such step.
Requires Windows 10 21H2 or later on x64 or ARM64.
Installs the app for all users and keeps it up to date.
Neither the installer nor the portable build is code-signed. When SmartScreen shows “Windows protected your PC”, choose More info, then Run anyway. Installing needs administrator permission, and the installer downloads WebView2 if it is missing; the portable build offers its download page instead.
Requires Ubuntu 22.04, Debian 12, Fedora 36 or a later distribution, on x86_64 or aarch64.
Downloads the AppImage for this machine, checks its sha256, places it in ~/Applications and starts it.
$ curl -fsSL https://github.com/ThinkWatchProject/ThinkWatch-Lite/releases/latest/download/install.sh | sh Download, allow it to run (chmod +x) and open it.
Needs fusermount3 from the fuse3 package, which most desktops include. Keep a downloaded AppImage in a folder the user can write to, such as ~/Applications, so that it can update itself. On GNOME the tray icon needs the AppIndicator extension.
Use, modification, and redistribution are permitted.
