proxy-llms

Run any model in t3's Codex and Claude tabs. A GPT model in the Claude tab, GLM or Kimi K3 in the Codex tab, plugged and unplugged on demand.

One local gateway (CLIProxyAPI) on 127.0.0.1:8317 speaks the OpenAI, Responses, and Anthropic formats at once and routes each request by model id. Register a model once, run it from either tab. Nothing touches t3's database, so unplugging leaves your setup as it was.

Setup

git clone https://github.com/0xhsn/proxy-llms ~/proxy-llms && ~/proxy-llms/setup.sh

It downloads the gateway, turns your existing Codex and Claude logins into gateway credentials, starts it, and routes the Codex tab through it. Idempotent, and it writes no secrets into the repo.

Commands

proxyctl status                 # gateway and plug state
proxyctl models [filter]        # what the gateway serves
proxyctl codex-plug             # Codex tab -> gateway (codex-unplug to revert)
t3-plugged                      # launch t3 with both tabs on the gateway; plain `t3` to unplug

proxyctl add-openai glm https://open.bigmodel.cn/api/paas/v4 glm-5 "$KEY"
proxyctl add-openai glm https://open.bigmodel.cn/api/paas/v4 glm-5 "$KEY" claude-fable-5
proxyctl remove glm

The Codex tab lists any model the gateway serves. The Claude tab shows only claude-* slugs, so give an external model a claude-* alias (second line above) to run it there. More provider base URLs in MODELS.md.

Notes

• Fable 5.1 needs claude-header-defaults.user-agent: claude-cli/2.1.255 in the config, or Anthropic rejects it. It ships set.
• The gateway is a background process, not a service. Run proxyctl start after a reboot.
• sk-local-7asan and the management key only guard 127.0.0.1. Change them in config/config.base.yaml before exposing the port.