FR: allow routing to original antropic and changing the model on the fly

[KIC]

Nice idea! I am wondering, since you already allow routing based in the model name to different providers, why not also allow antropic models to go antropic. I have a subscription there and I would switch away once the limit is hit, because, why should i not use my already payed subscription. It would also be cool to change the models via cli command instead of env variables which most likely even need a proxy server restart. This would additionally be a nice hack to solve the plan with model a and execute with model b problem which is not an option cc lets you pick in the promt.

[raine]

Interesting idea. I also (for now) have a subscription with Anthropic and what I do is toggle if proxy is used with claude-proxy-toggle, and restart claude with --continue.

% claude-proxy-toggle
proxy: off

Not saying this is the solution to you but I'll share some instructions on this in the README.

Restarting the proxy should not be necessary.

[KIC]

I see. In that sense one could also just create a sperate claude-with-proxy.sh put some exports and use the -c flag. Not very practical but sure works. Assuming env vars are used over settings.json.

Ps I was always annoyed that after plan mode, from the 3 options you were not able to change the model but you had to interrupt and kind of do everything manually. Now, if you say the proxy server picks up env variables at runtime (new values without restart), i should be able to export a new model in the same wrapper script right? Well it sure would not work within anthropic only but between proxy and anthropic. However , letting the proxy route to anthropic as well would be a great deal for both use cases.

[raine]

Now, if you say the proxy server picks up env variables at runtime (new values without restart), i should be able to export a new model in the same wrapper script right?

Claude Code in its requests sends model which controls to which provider the request is translated and sent to.

I'm really not sure (yet) if proxying to anthropic is a good idea - the proxy might have to start managing its own auth session unless Claude Code's can be reused somehow. Also, there's currently this hack in the proxy that in order to make skills that specify models (like sonnet) reusable, the model automatically translates sonnet to the model you are using (such as gpt-5.4). The existence of that hack conflicts with making it so that anthropic models are actually proxied to anthropic instead.

[KIC]

at the end of the day it is just directing traffic from one url to another. all the headers (auth), parameters and payload should stay unchanged. thus claude auth should work just fine. but maybe the selection criteria model name is not suitable for this case.

I try to use a script like claude-with-proxy.sh in the project for the moment and get the cheapest codex for a test month.

#!/bin/bash

if (timeout 2 bash -c "</dev/tcp/localhost/18765") 2>/dev/null; then
    echo "proxy is already running"
else
    echo "starting proxy"
    claude-code-proxy serve &
    sleep 1
fi

export DISABLE_AUTO_COMPACT=1
export ANTHROPIC_BASE_URL=http://localhost:18765
export ANTHROPIC_AUTH_TOKEN=unused
export ANTHROPIC_MODEL=${$1:-gpt-5.4}
export ANTHROPIC_SMALL_FAST_MODEL=${$2:-gpt-5.4-mini}
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

claude "$@"