⚠️ All API keys are saved only in your browser's localStorage — do not use on public computers. Some providers (e.g., Anthropic, OpenAI) require direct browser access; the corresponding headers are added automatically. For local models (Ollama), set OLLAMA_ORIGINS=* before launch to allow cross-origin requests.
Auto-detect tries /api/tags first — if found, uses Ollama native; otherwise falls back to OpenAI compatible.
Local models tend to slip into English once the context contains any English (history, attachments, code comments) — they start reasoning in English and drag the answer along. This appends a short language instruction to the end of the last user message (the position that actually works), and if the answer still comes back mixed, it appends a correction turn in the same context and takes the rewrite. Applies to local models only; other providers are never touched.
选择配色即可预览背景、按钮和消息气泡。点击保存后，会同步到当前账户的其他设备。
自动生成背景、按钮与消息配色
两种风格共用账号、API、模型和对话。此处可预览，点击保存后随账号同步。
管理常用提示词
从“常用提示词”选择后会显示为卡片，随下一条消息发送。
按问题给出充分说明和必要示例，也会优先遵守你在消息中的要求。
开启后回答会附上参考来源；关闭时不主动展示，追问来源仍可提供。随账户保存。
Shown above your avatar and messages. The avatar uses the first character.
Second-stage trimming on top of the message count above: the oldest messages are dropped until the estimated prompt fits. Estimation is rough (CJK about 1 char/token, other text about 4 chars/token). Set it somewhat below your model real context window.
Separate from the show-thinking switch below, which only controls display. For hybrid models such as Qwen3 this sends chat_template_kwargs.enable_thinking (vLLM / SGLang / llama.cpp --jinja) plus Ollama think - without it, hiding the thinking still pays for every thinking token.
Controls reasoning effort, not answer length. Detailed answers remain a separate choice.DeepSeek supports all three levels. Compatible GPT models use High for Maximum; other models may not support this control.
Ollama silently drops whatever does not fit its context window, and the first thing to go is the system prompt / scenario memory - with no error at all. Set this to the window your model actually supports (8192 / 32768 ...).
Empty fields are not sent at all. These apply to OpenAI-compatible and local models only. Qwen3 recommends thinking mode: temp 0.6 / top_p 0.95 / top_k 20 / min_p 0; non-thinking: temp 0.7 / top_p 0.8 / top_k 20. presence_penalty 0.5-1.5 is the usual cure when a small local model starts looping.
Vision tokens depend on image resolution, not file size — downscaling is what actually saves input tokens. Standard is visually near-lossless for most photos and screenshots.
Images/files are never carried forward as "memory": once a message with an attachment falls outside this many most-recent messages, only its text is resent — the attachment itself is dropped from the request. 1 = only the newest attachment is ever sent; 0 = never send attachments to the model at all.
For 🎬 scenario chats: when the live context exceeds the size below, older messages are automatically summarized into long-term memory.
Hard limit on the compressed long-term memory itself — it is hard-truncated to this length after every compression, so memory can never grow without bound. Memory only ever stores text; attachments are never included.
When off: panel is hidden; for Gemini 2.5, thinkingBudget:0 is requested to save tokens; DeepSeek-R1's <think> blocks are still stripped so they don't leak.
New scenario chat
Every past memory version (from auto-compression or manual edits) is kept here so you can roll back if something important got summarized away. Restoring loads it into the box above — click Save to apply.