Troubleshooting · API & Blank Replies
Why isn't SillyTavern responding? How to read the error and fix blank replies
SillyTavern stops responding when your upstream API rejects a call, times out, or hits its output cap, not because the UI crashed. Common culprits are a bad key (401), empty balance or rate limit (402/429), a mangled /v1 URL (404), or a reasoning model spending Max Response Length on hidden thinking tokens. On Android, Foreverse (free up to 3 cards) logs HTTP status and latency under Me → Request records.
Available for iPhone and iPad in selected regions outside mainland China. Not available on the mainland China App Store. Download help
What do SillyTavern error codes and blank bubbles actually mean?
SillyTavern is a local web client that forwards prompts to whatever backend you picked under API Connections. Before deleting presets or reinstalling Node.js, match what happened in the chat window to the HTTP status printed inside your SillyTavern terminal:
| Symptom in chat | Console / HTTP status | Root cause | Fastest fix |
|---|---|---|---|
| Typing indicator loops forever; never finishes | Socket timeout / ECONNREFUSED | Local backend (Ollama/KoboldCpp) isn't running, system VPN/proxy blocked the port, or phone OS suspended Termux in the background | Verify backend URL in browser, bypass proxy for localhost, or keep terminal awake |
| Toast: API returned an error (auth failed) | 401 Unauthorized / 403 Forbidden | Mistyped or revoked API key, trailing whitespace when pasting, or key restricted to a different region/project | Create a fresh key in provider console, re-paste without spaces, click Connect |
| Toast: quota or credit error mid-conversation | 402 Payment Required / 429 Too Many Requests | Prepaid balance hit $0.00, per-key spend limit reached, or free-model rate cap triggered (e.g., 20 RPM / 50 RPD on OpenRouter :free under 10 credits) | Check provider billing page, wait 60s for RPM reset, or switch off a :free model ID |
| Toast: endpoint or model not found | 404 Not Found | Custom base URL omitted /v1 or appended /chat/completions twice, or preset points to a retired model slug | End the custom URL at the provider's documented version prefix (commonly /v1) and re-pull the live /v1/models list |
| Generation stops instantly with an empty bubble | 200 OK (finish_reason: length or stop) | Reasoning model spent 100% of Max Response Length on hidden chain-of-thought, or a custom stop string clipped the first line | Raise Max Response Length (SillyTavern's docs suggest 1,024–4,096 for reasoning models), lower reasoning effort, or clear custom stop strings |
| Toast: API returned an error: OK | 200 OK containing an error JSON object | A third-party proxy or router wrapped an upstream capacity failure inside an HTTP 200 body | Disable Streaming for one turn to read raw console JSON; switch upstream provider |
Why do reasoning models finish with a completely blank bubble?
Green connection dot, funded balance, three seconds of generation—and then an empty speech bubble with zero red error toast. In 2026, that usually happens when you pair a thinking model with an older roleplay preset. SillyTavern's official Reasoning documentation spells out why: internal reasoning tokens count against your response token allowance even when hidden from chat (naming o3-mini and Gemini Thinking as examples), leaving little or nothing for the visible reply.
Classic roleplay presets routinely lock Max Response Length around 300 to 512 tokens to keep bots from writing walls of text. Give a reasoning model that tight cap and it can spend the whole allowance deliberating, hit finish_reason: length, bill you for those thinking tokens, and return an empty visible string, so no red error toast fires. SillyTavern's docs recommend raising Max Response Length into the 1,024–4,096 range for reasoning models, or switching to the provider's non-thinking chat slug if you want snappy turns.
Still getting empty bubbles on a standard non-thinking model? Check two quick suspects before blaming the API. Custom Stopping Strings such as {{char}}: or a bare newline will chop off the entire message if the model happens to open its turn with that exact prefix. And if Streaming is turned on, flip it off for one test turn: some failures get swallowed mid-stream and only print clearly in the server console when returned as a single JSON response.
How do you get SillyTavern generating again in five steps?
Walk through these five checks in order so you do not nuke a working preset or pay multiple times for half-finished generations:
- Uncheck Streaming and read the black terminal window
Turn off Streaming under AI Response Configuration, send a plain
Helloon a tiny test card, and look at the Node.js terminal (or your browser's DevTools Network tab). Toast banners truncate errors; the console prints the real HTTP status code and upstreamerror.message. - Check your Base URL ending and refresh /v1/models
Under Chat Completion API settings, custom OpenAI-compatible URLs should stop at the version root—like
https://openrouter.ai/api/v1orhttps://api.deepseek.com/v1—never/v1/chat/completions. Re-fetch the live model dropdown rather than pasting old slugs from Reddit; DeepSeek's pricing docs, for instance, now point users todeepseek-flashand note that legacydeepseek-v4-flashroutes to newer weights. - Verify rate caps and wallet balance on :free or router keys
According to OpenRouter's documented limits (checked 2026-11-15),
:freevariants cap accounts under 10 lifetime credits at 20 requests per minute and 50 requests per UTC day (jumping to 1,000/day once you buy 10+ credits), and a negative credit balance can cause errors even on free slugs. Mashing Regenerate five times during a 429 bottleneck only extends your cooldown. - Bump Max Response Length and isolate heavy cards or lorebooks
Push Max Response Length into the 1,024–4,096 token range SillyTavern's docs mention whenever reasoning is active, and make sure Context Size fits inside the model's real window. In our census of 19,776 public cards, 90% of the top 1,000 wrapped their greeting in HTML—downloaded cards can be much heavier than they look. Toggle World Info off for one turn to rule out context overflow or runaway recursive lorebooks.
- Turn Streaming and World Info back on one piece at a time
Once that minimal test card replies with a clean 200, re-enable Streaming first and World Info second. If things only freeze on one specific downloaded card, drop that file into our free browser Character Card Checkup to spot broken macros or oversized permanent fields.
How can you read upstream API errors on a phone without Termux?
Running desktop SillyTavern on a phone—whether inside Android Termux or in a browser tab pointed at a home PC—adds a layer of mobile-specific breakage that has nothing to do with your API key. Lock your screen or jump over to reply to a text mid-turn, and the phone can suspend the background connection, while the HTTP codes you need sit in a terminal window that is awkward to read on a small screen.
Our team builds Foreverse, a native Android (Google Play and direct APK) and iPhone/iPad (App Store, not available in mainland China) app for Character Card v2/v3 roleplay, AI companions, and branching novels, so evaluate our pitch with that in mind. Foreverse is free to install and use with your own API keys for up to 3 stored character cards, 3 books, and 3 companions (a one-time $19.99 Pro purchase lifts library caps with no monthly fee). Rather than making you alt-tab into a terminal window, every BYOK and official-channel call writes to an on-device log under Me → Request records.
Opening Request records shows HTTP status, latency, token counts, and channel for each attempt, while Test connection under Me → AI Models & Services prints the status code and the provider's raw error message before you enter a chat. Keep one billing reality in mind: a failed display on your phone does not guarantee the upstream server generated zero tokens, so check your provider's usage dashboard before hitting retry on a huge prompt. Over 60 preset providers come pre-filled with their exact versioned API host, Import from upstream /v1/models pulls live model IDs straight from the server, and Manage multi-key & rotation lets you pool backup keys and rotate or fail over between them when one is throttled. Our walkthrough is Configure BYOK and Fix 401/404/429 Errors.
When Foreverse won't solve your SillyTavern problem
- You run local GGUF or EXL2 weights on a desktop GPU via Ollama, KoboldCpp, or TabbyAPI: freezes there come from VRAM exhaustion, port conflicts, or firewall rules that only your PC backend logs can diagnose.
- Your setup depends on desktop SillyTavern server plugins, Python Extras modules, or custom Node.js middleware rather than portable Character Card v2/v3 files, lorebooks, and card-level HTML/regex.
- You want one browser session synced live between your PC and phone: Foreverse runs natively on Android and iOS with local storage and offers no web chat client or cloud sync.
FAQ
What does 'API returned an error: OK' mean in SillyTavern?
Your API gateway or reverse proxy sent back HTTP status 200 (whose standard status text is 'OK') while stuffing an error JSON object inside the body instead of chat choices. Turn off Streaming for one message and look at the SillyTavern terminal window to read the actual upstream error.
Why do DeepSeek or Gemini thinking models return blank messages in SillyTavern?
As SillyTavern's reasoning docs explain, hidden chain-of-thought tokens still count toward Max Response Length. When a preset caps responses at a few hundred tokens, the model can spend every token thinking and stop before writing visible dialogue. Raise Max Response Length (SillyTavern suggests 1,024–4,096 for reasoning models) or switch to a non-thinking model ID.
Why did my OpenRouter :free model stop working mid-chat?
OpenRouter caps :free slugs at 20 requests per minute and 50 requests per UTC day on accounts with fewer than 10 lifetime credits purchased (or 1,000/day with 10+ credits), and busy free upstream hosts can throw 429 errors of their own. Wait sixty seconds, top up 10 credits once, or switch to the paid version of that model.
Can a single character card or lorebook make SillyTavern hang?
Yes. Recursive lorebook entries without cooldowns or delay-until-recursion guards can balloon a prompt past the model's context ceiling or make every request much larger and slower. Switch World Info off for a single test message or drop the card into /labs/card-checkup to check its token footprint.
Does hitting Regenerate after a failed turn charge my API key twice?
Immediate 401, 403, 404, or 429 errors are rejected before generation starts, so providers normally charge nothing. If your network dropped mid-stream after the model began generating tokens, the upstream lab may still bill whatever tokens it produced—check your provider's usage log or Foreverse's Request records before retrying long contexts.
Sources
- SillyTavern Docs — Reasoning & Max Response Length (checked 2026-11-15)
- SillyTavern Docs — Frequently Asked Questions (checked 2026-11-15)
- SillyTavern Docs — Chat Completions & OpenAI-Compatible API Setup (checked 2026-11-15)
- OpenRouter Docs — Credit Limits & Free-Model Rate Limits (checked 2026-11-15)