What is a SillyTavern proxy, and how do you pick a safe one?

What does "proxy" mean in SillyTavern?

In AI roleplay communities, a proxy is simply an API endpoint URL that sits between your chat frontend (such as SillyTavern or JanitorAI) and an upstream AI model provider such as Anthropic, Google, or DeepSeek. When you hit send, your frontend sends the assembled prompt to that URL, the service forwards the request to the model, and the reply streams back to your screen.

Players started calling almost every third-party API URL a "proxy" because of how the apps themselves label their settings. SillyTavern's native Claude and Google AI Studio panels put third-party URLs under a heading called Reverse Proxy, and JanitorAI puts its bring-your-own-API fields under a tab called Proxy.

At the same time, if you spend five minutes on community forums or Discord servers, you will see experienced players tell newcomers to stay away from proxies. That caution is well-earned. In roleplay circles, the word "proxy" is heavily associated with free community links passed around on anonymous paste pages—endpoints that run on leaked or stolen API keys, quietly log people's private chats, swap expensive models for cheap ones, and vanish without warning.

To sort out the confusion, you have to separate the UI label from how the service behind it operates:

• A free or shared-key proxy gives you a single public URL (or a password shared in a group chat) without clear billing or an identifiable operator, usually funded by abused trial accounts or compromised keys.
• A paid API proxy (or API gateway) works like a standard cloud API service: you sign in to your own account, generate a private API key, and pay per token for access to named upstream models.

In the rest of this guide, we use proxy wherever your app's own screen uses that word (Reverse Proxy, Proxy URL) and API as the normal noun for the service you connect.

Which kind of proxy do you need?

Before you edit any configuration files, make sure the guide or forum post you are reading is talking about the same kind of proxy you actually want. Across SillyTavern documentation and community threads, people use the word "proxy" for four completely different tools:

Where does the proxy URL go?

Every roleplay app asks for the same three things—an endpoint URL, an API key, and a model ID—but different apps expect different parts of the URL path. Below is exactly where those fields live in each client, using TavernAPI and our five model IDs (claude-opus-5-5, claude-sonnet-5-5, gemini-4-argon, gemini-3.8-flash, and deepseek-v4-pro) as worked examples.

SillyTavern: Chat Completion → Custom (OpenAI-compatible)

Connecting through Custom (OpenAI-compatible) is the cleanest option in SillyTavern because a single endpoint and key let you switch between Claude, Gemini, and DeepSeek from one dropdown:

1. Click the API Connections icon (the electric plug on the top bar).
2. Set API to Chat Completion.
3. Set Chat Completion Source to Custom (OpenAI-compatible).
4. Fill in the connection fields:
   • Custom Endpoint (Base URL): https://tavernapi.com/v1
   • Custom API Key: Your sk-tav-... key (created in the user console; copy it right away because it is shown only once).
5. Click Connect, open the Available Models dropdown, and select gemini-4-argon, claude-opus-5-5, gemini-3.8-flash, deepseek-v4-pro, or claude-sonnet-5-5.

Do not add /chat/completions to the end of Custom Endpoint (Base URL)—SillyTavern adds that suffix automatically when sending a message.

SillyTavern: Claude and Google AI Studio Reverse Proxy fields

If you use a local preset designed specifically for SillyTavern's native Claude or Google AI Studio source—or if you want source-specific UI controls for Claude or Gemini—you can use SillyTavern's Reverse Proxy section instead of Custom (OpenAI-compatible).

1. Open API Connections (the plug icon) and keep API set to Chat Completion.
2. Set Chat Completion Source to either Claude or Google AI Studio.
3. Click Reverse Proxy to expand the proxy fields and enter the matching URL:
   • For the Claude source, set Proxy Server URL to https://tavernapi.com/v1
   • For the Google AI Studio source, set Proxy Server URL to https://tavernapi.com (no /v1 at the end, because SillyTavern appends Google's versioned path itself)
4. In Proxy Password, paste your sk-tav-... API key. (SillyTavern calls it a password, but it carries your API key.)
5. Click Connect and select claude-opus-5-5 or claude-sonnet-5-5 under Claude, or gemini-4-argon or gemini-3.8-flash under Google AI Studio.

For a full walkthrough of both connection methods and preset imports, see our SillyTavern setup guide. If you ever see a 402 error when sending a prompt, check how to fix 402 Insufficient Balance.

JanitorAI: API Settings → Proxy

JanitorAI is a hosted character site with its own built-in model, but it also lets you route chats through an external OpenAI-compatible API:

1. Open any character chat on JanitorAI, click API Settings at the top of the screen, and select the Proxy tab.
2. Fill in the three required fields:
   • Proxy URL: https://tavernapi.com/v1/chat/completions
   • API Key: Your sk-tav-... key
   • Model: Type the exact model ID you want to use, such as gemini-4-argon, claude-opus-5-5, gemini-3.8-flash, deepseek-v4-pro, or claude-sonnet-5-5.
3. Save your settings and start chatting.

RisuAI, Agnaistic, and other OpenAI-compatible apps

Other roleplay frontends such as RisuAI and Agnaistic (Agnai) also accept any custom OpenAI-compatible endpoint (see our comparison of the best AI roleplay tools for how each app works).

To see which URL format a client expects, look at the grey placeholder example inside its URL box:

• If the placeholder ends in /v1, enter https://tavernapi.com/v1.
• If the placeholder ends in /v1/chat/completions, enter https://tavernapi.com/v1/chat/completions.

Then paste your sk-tav-... API key and enter one of the five model IDs (gemini-4-argon, claude-opus-5-5, gemini-3.8-flash, deepseek-v4-pro, or claude-sonnet-5-5).

Why are free proxies risky?

Running large language models like Claude and Gemini costs money on every single turn—especially in roleplay, where your app re-sends thousands of tokens of character definitions and chat history with every reply. Both SillyTavern's official documentation and Chub's documentation warn players about the risks of using untrusted third-party proxies.

When someone hands out a free or suspiciously cheap proxy link with no clear business model behind it, you are taking on four concrete risks:

1. Leaked or stolen keys get cut off without warning: Free proxies rarely pay for their own upstream compute. In practice, their key pools are often scraped from public code repositories, opened with stolen billing credentials, or cycled across abused cloud trial accounts. As soon as the upstream provider detects the unauthorized traffic, the keys are revoked and your proxy dies mid-scene.
2. The operator can log and read your private chats: Every message you send through a proxy passes through the operator's server in readable text so it can be forwarded to the model. An anonymous operator with no privacy policy can log your prompts, character cards, user personas, and IP address, and you have no way to know where those logs end up.
3. The model behind the endpoint may not be the one advertised: Because the proxy server controls the outgoing request, a dishonest operator can take a request labeled for a flagship Claude or Gemini model and silently route it to a much cheaper, smaller model instead. If a supposed top-tier model suddenly forgets basic scene logic or writes flat, repetitive sentences, the endpoint may be swapping models behind the scenes.
4. No support, no accountability, and no refunds: Informal proxies and "donation-tier" Discord servers can disappear overnight when their key supply runs dry. There are no published terms, no support channels, and no way to get your money back if you paid for a shared slot.

How do you check a proxy before you pay?

You do not need to be a network engineer to separate a legitimate API service from a sketchy operation. Run any provider through this seven-point checklist before you top up:

1. It names its upstream providers per model: A trustworthy service states clearly which official cloud or API channel serves each model (such as AWS, Google Cloud, Google AI Studio, or the official DeepSeek API), rather than hiding behind vague claims of "private routes."
2. You pay for your own usage with dated, public prices: Look for a public pricing table with a visible "last updated" date where you are billed per token (or through clearly explained credit packs) for your own personal API key.
3. A published privacy policy states what is stored and what is not: Check for a real privacy policy page that explicitly distinguishes billing metadata (token counts, model name, timestamp, and cost) from your private prompt text and character cards.
4. There is a reachable operator and published terms: Make sure the site publishes Terms of Service and gives you more than one working channel—such as email, Telegram, or Discord—to reach a real person if something goes wrong with a top-up.
5. The model you get matches the model you picked: Test the endpoint with a character card or a tricky multi-character scene you already know well to confirm that the prose style, reasoning depth, and response speed match the genuine model.
6. The price is plausible, not "free unlimited Claude": Enterprise volume discounts and server-side prompt caching can legitimately cut per-token costs, whereas anyone offering flat-rate "free unlimited Claude" is cutting corners you do not want near your account.
7. You can start with a small top-up and check the usage log: Never put a large balance into an unfamiliar service on day one; start with a free trial or the smallest top-up, send a few test replies, and verify that the usage log in the console matches the published per-token rates.

Where does TavernAPI stand?

To keep this guide fair, let's apply that exact same seven-point checklist to TavernAPI, along with the honest trade-offs of our service:

• Named official upstreams: We carry five 1M-context models chosen specifically for roleplay (read our hands-on model rankings). claude-opus-5-5 and claude-sonnet-5-5 run through AWS and Google Cloud, gemini-4-argon and gemini-3.8-flash run through Google AI Studio, and deepseek-v4-pro runs through the official DeepSeek API.
• Dated per-token prices in USD: All input, output, and cache rates are published in USD per million tokens on our pricing page. We also apply automatic server-side prompt caching on every connection to lower long-chat input costs and display your exact dollar savings in the user console.
• No chat content storage: As detailed in our Privacy Policy, TavernAPI does not store your message content, character cards, lorebooks, or AI replies. We log only token counts, request time, model ID, and cost for billing.
• Reachable operator and clear terms: We publish full Terms of Service and offer direct support via Telegram, Discord, or email on our contact page.
• Test before you pay: New accounts receive a $0.30 free trial on gemini-4-argon and gemini-3.8-flash so you can verify model quality and inspect per-request token logs in the console before spending a cent. For SillyTavern players, we also provide optional storytelling presets that combine local toggles with server-expanded core instructions.

Our honest limits: TavernAPI is not free once you use up your $0.30 trial credit. Because every request goes through official upstream providers, the providers' own safety filters still apply (though we never ban your account just because an upstream model refused a prompt). Finally, we only carry the five Claude, Gemini, and DeepSeek models listed above—if you want OpenAI GPT models or dozens of small open-weight finetunes, another gateway will fit you better.