ThinkWatch-Lite puts a local gateway between Claude Code and the model, and reached 426 GitHub stars in 20 days
ThinkWatch-Lite is an MIT-licensed local gateway that sits between coding assistants and the model API, letting a team switch upstreams without editing each client, keep real API keys out of relay services, and block the tool calls that try to steal them. The repository was created on 12 September and reached 426 GitHub stars by 2 October.
Why it mattersA team using several AI coding assistants can rotate a key, switch providers, or block a prompt-injection payload in one place, instead of opening every tool and finding the right config file.
Rotating the API key under a coding assistant usually means opening every tool that uses it and editing a config file by hand. ThinkWatch-Lite, a new open-source local gateway published by the ThinkWatchProject organisation, moves that one job to one place. It sits between coding assistants and the model API, so Claude Code, Claude Desktop, Codex and other clients point at the gateway and the gateway decides which upstream the request actually reaches. The repository was created on 12 September 2026 and reached 426 GitHub stars by 2 October, from the stargazers_count field on the GitHub API.
One point of control for several coding assistants
The README states four features: switch upstreams without touching the clients, keep API keys from relay services, block malicious tool calls, and trace every request. The clients are pointed at the gateway once; after that the user picks the upstream from the gateway and the clients see no change. The project ships with ThinkWatch Core embedded, under an MIT licence, and lists installers for macOS 12 or later on Apple silicon, Windows 10 21H2 on x64 and ARM64, and Linux on x86_64 and aarch64.
Two inspection steps that read both sides of a request
The outbound redaction step rewrites API keys, private keys, JWTs, connection-string passwords, Chinese resident ID numbers and bank card numbers to placeholders before a request leaves the machine. The inspection step, on the way back, looks at tool-call answers and blocks the ones that try to download and run code, send out environment variables or credential files, read private keys, install a startup item, or schedule a job. Both lists come from the project's README. These two steps address the same risk from each side: a prompt-injection payload reaching one of these models produces an answer that asks the client to perform the harm, and the client reads that answer before the user ever sees it. The gateway reads the request on the way out and the answer on the way back, which is the only place both halves can be checked.
The full request log, searchable and replayable
The request history is searchable and holds the full request and response text, the rule that matched, the upstreams the gateway tried, the API-format conversions it made, and the cost it calculated. A finished request can be replayed against another upstream and the two outputs compared side by side. The repository description also lists MCP server and client configuration scanning, routing rules with failover, and support for several API formats so one request can be rewritten for a different provider.
The gateway runs as an unsigned desktop app alongside the clients, so it is a local tool that each developer installs on the same machine as their coding assistants. A team sharing one gateway on a server would need something different. For a working developer with Claude Code and Codex on the same laptop, the local setup is already the common one. The gain is a single place to change the upstream, read what each tool actually sent, and refuse the one dangerous answer; most teams today keep all three in a separate config file inside each tool.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
