ChatGPT Next Web

Abstracts multiple LLM providers (OpenAI GPT-4, Anthropic Claude, custom endpoints) behind a unified chat API, allowing users to switch providers and models without UI changes. Implements provider-agnostic message formatting, token counting, and streaming response handling through a pluggable backend architecture that normalizes API differences across OpenAI, Anthropic, and custom HTTP endpoints.

Automatically summarizes older conversation turns into compressed context when approaching token limits, preserving semantic meaning while reducing token consumption. Uses a recursive summarization strategy that condenses multi-turn dialogues into concise summaries, allowing long conversations to continue without hitting model context windows or incurring excessive API costs.

Tracks token consumption for each message and conversation, displaying cumulative token counts and estimated API costs based on current pricing. Uses model-specific token counting (via tiktoken for OpenAI, manual counting for other providers) to estimate costs before sending requests, helping users understand API expenses and optimize prompt length.

Implements a responsive design that adapts to mobile, tablet, and desktop viewports, with touch-optimized buttons, swipe gestures for navigation, and mobile-specific layouts. Uses CSS media queries and touch event handlers to provide a native app-like experience on smartphones without requiring a separate mobile application.

Streams LLM responses token-by-token to the UI as they arrive from the provider, rendering each token incrementally rather than waiting for the complete response. Uses Server-Sent Events (SSE) or WebSocket connections to receive streaming data, with real-time DOM updates to display tokens as they arrive, providing immediate feedback and perceived responsiveness.

Allows users to branch conversations at any point, creating alternative response paths without losing the original conversation. Each branch maintains independent message history, and users can compare branches side-by-side or merge insights back into the main conversation. Implements a tree-based conversation structure where each message can have multiple child branches.

Provides a built-in library of pre-written prompt templates with parameterized variables (e.g., {{topic}}, {{tone}}) that users can customize and execute. Templates are stored locally or fetched from a remote repository, parsed for variable placeholders, and rendered with user-provided values before sending to the LLM, enabling rapid prompt reuse without manual editing.

Parses LLM responses for markdown syntax and renders formatted text, code blocks, tables, and lists in the chat UI. Uses a markdown parser (likely remark or markdown-it) with syntax highlighting for 50+ programming languages via Prism.js or highlight.js, enabling readable code snippets and formatted content directly in conversations.

Provides a pre-configured Vercel deployment template that automatically forks the repository, configures environment variables (API keys, model selections), and deploys the web UI to Vercel's serverless infrastructure with a single click. Handles build configuration, environment setup, and domain assignment without requiring manual CLI commands or infrastructure knowledge.

Implements PWA standards (service workers, manifest.json, offline caching) enabling the web app to be installed as a desktop or mobile application and function offline for previously loaded conversations. Service workers cache static assets and conversation history, allowing users to read past messages and compose new ones without network connectivity, with automatic sync when connection is restored.

Allows users to export conversations as JSON, Markdown, or PDF files, preserving message history, metadata, and formatting. Conversely, users can import previously exported conversations to restore full chat history including system prompts and model selections. Export/import uses standard file formats enabling portability across different chat applications and long-term archival.

Allows users to define custom system prompts that shape model behavior for the entire conversation, with pre-built role templates (e.g., 'Code Reviewer', 'Creative Writer', 'Technical Explainer'). System prompts are stored per conversation and applied to all subsequent messages, enabling consistent persona and behavior without repeating instructions in every message.

Implements full-text search across all conversations stored in the browser, allowing users to find specific messages, topics, or responses by keyword. Search results are ranked by relevance and filtered by date range, model used, or conversation topic, enabling rapid retrieval of past insights without manual scrolling through conversation history.

Provides UI translations for 10+ languages (English, Chinese, Japanese, Spanish, French, etc.) with dynamic language switching without page reload. Uses a client-side i18n library (likely i18next or similar) to load language packs and apply translations to all UI elements, buttons, and messages in real-time.