SillyTavern is the power user's front end for AI roleplay: a free, open-source interface under the AGPL that you install on your own machine, load characters and lorebooks into, and point at whatever language model you like. Its documentation says it plainly: SillyTavern is only an interface, so you need access to an LLM backend to provide inference. That backend can be a model running on your own GPU through KoboldCpp, llama.cpp or Ollama, or a cloud API such as OpenAI, Claude, Google AI Studio, Mistral or OpenRouter.
Because SillyTavern connects to so many things, the word "proxy" has come to mean three different things inside it, and a person searching for a SillyTavern proxy could want any of them. One is a field in the settings that takes a language-model address. One is a way of exposing SillyTavern itself to other devices. Only the third is a proxy in the sense we sell, an address your traffic passes through, and it is the one most guides never mention because it lives in a config file rather than the interface. This page sorts the three, then explains when the network kind matters and when it changes nothing.
The three proxies of SillyTavern
| What people call it | Where it lives | What it changes |
|---|---|---|
| The Reverse Proxy field | Chat Completion settings: Proxy Server URL, Proxy Password | Which server answers as the language model |
| An HTTP reverse proxy | The documentation's reverse-proxying guide (nginx, Caddy) | How other devices reach your SillyTavern |
| requestProxy | config.yaml: enabled, url, bypass | The IP address your API calls come from |
The documentation itself draws the first line. Its reverse-proxying guide opens by stating that the section does not refer to OpenAI or Claude reverse proxies and refers exclusively to HTTP and HTTPS reverse proxies, which is the second row. The first row is what the roleplay community means, and the third is ours.
Meaning one: the Reverse Proxy field is a language-model address
Switch the API type to Chat Completion and the settings show a Reverse Proxy block: proxy presets with saved addresses and passwords, a Proxy Name, a Proxy Server URL described as an alternative server URL, and a Proxy Password used in place of an API key. Next to it sits the Custom (OpenAI-compatible) source with a Custom Endpoint (Base URL) and an optional Custom API Key, with the interface's own hint that if the endpoint does not work you should try adding /v1 at the end, because the /chat/completions suffix is added automatically.
Everything in that block is about which server writes the replies. Point it at OpenRouter and OpenRouter's models answer. Point it at Chub's inference API, which Chub's documentation says can be used in the same way as an OpenAI reverse proxy, and Chub's Mars models answer. Point it at a community "free proxy" posted in a Discord and a stranger's server answers, on a key that is usually not theirs to use. This is exactly the setup that Janitor AI users mean by a proxy, and we cover the free versions of it in best free SillyTavern proxy and, for the neighbouring app, best free Janitor AI proxy.
Two warnings sit right in the product. SillyTavern's interface states that using a proxy you are not running yourself is a risk to your data privacy, and that any support requests will be refused if you are using one. Chub's documentation for the same feature goes further: reverse proxies are created by unauthorized third parties who gain access to your data and IP address, may enable logging, leak your information, and use stolen API keys which you have no control over. Nothing in that block changes your IP address, and a network proxy does nothing to it. Keep the two ideas apart and the rest of this page gets simple.
Meaning two: exposing SillyTavern itself
The second meaning is infrastructure. SillyTavern runs a local web server, and people who want to reach it from a phone or another machine put nginx or Caddy in front of it as an HTTP reverse proxy, with a domain and a certificate. The documentation's guide covers that, along with the whitelist and basic-authentication settings that keep strangers out. It has nothing to do with the model or with your outbound address, and it is not what we sell, so we leave it there.
Meaning three: requestProxy, the network proxy for your API calls
The third meaning is the one that matches the word everywhere else on the internet. SillyTavern's config.yaml has a requestProxy block: enabled, off by default, turns it on; url takes the proxy server address, with the documentation's example in the form socks5://username:password@example.com:1080; and bypass lists hosts that should skip the proxy. Once enabled, SillyTavern sends its outbound requests, the calls to OpenAI, Anthropic, Google, OpenRouter or any custom endpoint, through that address, so the provider sees the proxy's IP instead of yours. The documentation adds one caveat: request proxying conflicts with the private-address whitelist feature, so if connections fail after enabling it, that setting is the first thing to check.
# config.yaml, SillyTavern release branch
requestProxy:
enabled: true
url: "http://USER:PASS@gateway.hproxy.com:PORT"
bypass:
- localhost
- 127.0.0.1
Restart SillyTavern after the change, reconnect the API, and confirm the exit with our proxy checker before you rely on it. Local backends such as KoboldCpp on the same machine stay in the bypass list, because there is no reason to send localhost traffic anywhere.
When the network proxy matters, and when it changes nothing
The cases where requestProxy is the right tool are narrow and real.
A provider that does not serve your country. The cloud APIs behind SillyTavern each publish where they are offered, and none of them serves everywhere. Anthropic lists the countries where Claude and its API are available and excludes places such as mainland China, Hong Kong, Macau, Russia and Iran, as we detail in proxies for Claude AI. OpenAI maintains a separate supported-countries page for its API. Google's Gemini free tier is not offered in every country, which is why it comes up so often in Janitor AI threads. DeepSeek's Open Platform terms bar use in comprehensively sanctioned countries. A request from an unsupported country is refused before any quota is counted, and a static residential address in a supported country is what changes the answer. It is also a decision to read against the provider's terms, which say where they choose to sell, so read them first.
A network that filters the API hosts. Universities, offices and some national networks block api.openai.com or api.anthropic.com outright. Routing through a proxy the filter does not recognise makes the calls go through, the same mechanism as any blocked site.
A consistent exit for an API account. API accounts carry a billing country and a usage history, and a key that has only ever called from Germany and starts calling from Vietnam draws exactly the security review you were avoiding. A fixed exit, held permanently, keeps the account looking like itself.
And the cases where it changes nothing, which is most of them.
Rate limits and quotas. A 429 from any provider follows the key in the request header. The same key through a different address is the same 429. We set out the whole argument in LLM API rate limits and proxies, and it applies to every backend SillyTavern can connect.
Content moderation. A provider's filters run on the request, not on where it came from.
A local model. If your backend is KoboldCpp or Ollama on your own hardware, there is no outbound call to route and no IP to change. requestProxy is irrelevant, and that is most of what makes local models attractive.
A static residential IP moves
A provider that does not serve your country
the request is refused before any quota counts
A network that filters the API hosts
office, campus, national filter
The country an API account appears from
one fixed exit, held for good
It never touches
429s and quotas
follow the API key
Content filters
run on the request, not the address
The Reverse Proxy field
a model address, not an IP
Local backends
nothing leaves the machine
Which type, and why static
Because the point is a consistent country for an API account, the type is one static residential or ISP address held permanently, never a rotating pool. Rotation solves per-IP rate limits on websites that meter by address; API providers meter by key, so rotation buys nothing there and a moving address is itself a signal. An ISP address is registered under a consumer ISP, so it reads as a home connection, and it is hosted on always-on hardware, so it does not drop mid-session the way a real home line might. A static residential session on a real home line is the alternative when a particular country matters. Datacenter addresses are the profile most providers already associate with automated abuse, and free proxies fail for the reason SillyTavern's own warning gives about reverse proxies: a stranger's machine in the middle of your traffic is a risk to your data, and the API key travels in every request. Our explainer on rotating versus static residential proxies covers the split in full.
Where HProxy fits
requestProxy takes one URL, and the URL that fits is a dedicated ISP proxy: a static IP with an ISP's name on it, yours alone, $2.70/IP a month. If the country matters more than the uptime guarantee, a static residential session from $0.44/GB goes in the same field. Chat completions weigh almost nothing, so the pricing, metered as you go with no subscription, is effectively the price of the address. Paste it, restart, check the exit with the proxy checker, and leave it alone.
If what you were actually looking for is a free model to put in the Reverse Proxy field, you do not need us at all; best free SillyTavern proxy has the backends, the exact values to paste, and the errors you will meet. The network proxy is for the narrower problem of where your API calls come from, and for that, one static address settles it.
Sources
- SillyTavern documentation, introduction: a locally installed interface that needs an LLM backend; AGPL-3.0.
- SillyTavern documentation, config.yaml reference: the requestProxy block (enabled, url, bypass) and its conflict with the private-address whitelist.
- SillyTavern documentation, OpenAI and Chat Completion: the Custom (OpenAI-compatible) source, the custom endpoint, and the /v1 hint; and the reverse proxying guide, which states that it does not refer to OpenAI or Claude reverse proxies.
- SillyTavern release-branch interface source, public/index.html: the Reverse Proxy block (Proxy Server URL, Proxy Password), the Custom Endpoint (Base URL) field, and the warning that support is refused to users of third-party proxies.
- Chub documentation, API connections: the warning about third-party reverse proxies, and usage with third-party UIs for the Mars and Mercury APIs used as an OpenAI-style reverse proxy.
- Anthropic, supported countries and regions; OpenAI, API supported countries; DeepSeek, Open Platform terms of service.