Sage Router

Local-first AI model routing for serious agents. One endpoint. Any provider. The router figures out the rest.

This package implements a capable local model router, but it includes several convenience features that touch sensitive local state and can expose your router to the Internet. Before installing or running it, consider the following: - Review grok_sso_proxy.py carefully and decide whether you want the Grok SSO feature: it will read browser cookie DBs by default and can call the Bitwarden CLI (bw) to unlock and fetch stored items. If you don't want that, do not set GROK_SSO_* env vars, do not run the proxy, and do not mount your browser profile or ~/.dario into the container. - If you deploy the Cloudflare Worker + connector or follow the Cloud Run deployment, be aware you can (if misconfigured) expose your local router publicly. Never deploy provider keys, cookies, or .env files to public Cloud Run or other public clouds; follow the repo's own caution but double-check the commands (the provided Cloud Run script uses --allow-unauthenticated by default). - Limit mounts: avoid mounting your real ~/.openclaw, ~/.dario, browser profiles, or ~/.ssh into an unvetted image. Instead run in an isolated VM/container and mount only the minimal config you trust. - Disable autostart features you don't need: set SAGE_ROUTER_DARIO_AUTOSTART=0 and SAGE_ROUTER_BUNDLED_OLLAMA_AUTOSTART=0 if you don't want bundled daemons started. - Audit env vars and secrets: the repo references analytics and Supabase/Firestore tokens. If you enable analytics mirroring, understand what telemetry is sent and where. - If you lack time or expertise to audit the code, treat this as untrusted: run it in a disposable environment, or avoid enabling the Grok SSO features and remote tunnel. Ask the author to explicitly list all env vars and side-effects (cookie reading, Bitwarden use, autostart behaviors) in the top-level metadata before using on a machine with sensitive data.

Type: OpenClaw Skill Name: sage-router Version: 3.26.23 The sage-router bundle is a complex AI model routing gateway with several high-risk capabilities. Most notably, 'grok_sso_proxy.py' is designed to read and decrypt browser cookies (specifically from Brave/Chrome) and interact with the Bitwarden CLI ('bw') to retrieve credentials. Furthermore, the bundle includes a remote access tunnel implementation ('edge/cloudflare-ai-tunnel/worker.js' and 'scripts/sage_tunnel_connector.mjs') that allows remote execution of chat completions via WebSockets. While these features are aligned with the stated goal of providing a 'local-first' gateway for various AI providers, the combination of browser data access, credential manager interaction, and remote tunneling represents a broad and risky attack surface without clear proof of malicious intent.