The AI tools directory collects 127 AI products and platforms that people are genuinely using right now, in nine groups: chat assistants (ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen, Doubao…), writing and office work (polish, translation, slides, meeting transcription), image and video generation (Midjourney, Firefly, Jimeng, Kling, ComfyUI…), voice and music (ElevenLabs, Suno, Whisper), coding assistants (Cursor, Copilot, Claude Code, Lingma, Continue), agents and workflows (Dify, Coze, n8n, LangChain, MCP), local running (Ollama, LM Studio, llama.cpp), benchmarks and leaderboards (LMArena, Artificial Analysis, OpenCompass) and API and model platforms (OpenAI, DeepSeek, SiliconFlow, OpenRouter…).

For users in China the first question about an AI tool is not “which one is strongest” but “can I sign up, does it cost money, does it need a foreign phone number, and where does my data go”. Every entry therefore carries four labels: Free (a usable free tier exists), Works in China (opens and works over a direct connection, labelled conservatively), Foreign phone (sign-up requires verifying a number from a supported region) and API (can be wired into your own program or into this site's tools), plus a Local / on-site tag for models that run on your own machine or inside toolun. Five toggles at the top of the page combine into filters — “Free + Works in China + Has API”, for instance, gives you the providers you can paste straight into this site's AI Chat.

The site's own AI tools sit in a group of their own at the front. They run in BYOK (Bring Your Own Key) mode: you enter your own API key in the settings, the request goes straight from your browser to the provider you picked, and toolun's servers never touch the content. Cutout, OCR and face blur go further and use purely local models, so the images never leave the browser.

1. 01Start from your task— for chatting, open “Chat assistants”; for drawing, “Image and video generation”; to run models yourself, “Local running”; to wire up an API, the last group.
2. 02Filter out what you cannot use— tick “Works in China” and “No foreign phone number” to see what can be registered and used from China directly; tick “Free” to drop the paid-only products.
3. 03Search— type a product name, a use or a tag (“transcription”, “text to image”, “open source model”, “knowledge base”).
4. 04Read the review and the badges— the one-line review states the strength and the limit (“free tier is low resolution only”, “needs a phone number from a supported region”, “no free trial”), and the badges carry the labels.
5. 05Wire it into this site— among the “API and model platforms”, the providers marked as working in China (DeepSeek, SiliconFlow, Model Studio, Ark, Zhipu, Moonshot) hand you a key that goes into the settings of AI Chat; Ollama also accepts a local address.
6. 06Bookmark— the star pins an entry to the top of the page (stored on this device).

Tick “Free” + “Works in China” + “Has API” and the “API and model platforms” group gives you:

Wiring DeepSeek into this site's AI Chat settings:

DeepSeek Platform        OpenAI-compatible, very cheap, local payment; paste straight into AI Chat.   Free · Works in China · API
SiliconFlow              Open-source model APIs from China, credit on sign-up; Qwen / DeepSeek / FLUX.   Free · Works in China · API
Alibaba Cloud Model Studio   APIs for Qwen and third-party models, free quota for new users.   Free · Works in China · API
Volcano Engine Ark       Doubao model APIs, low price and high concurrency.   Free · Works in China · API
Zhipu BigModel           GLM series APIs, free models available.   Free · Works in China · API
Moonshot Open Platform   Kimi model APIs, long context.   Free · Works in China · API

Endpoint       https://api.deepseek.com/v1
API Key        sk-… (created on the DeepSeek Platform)
Model          deepseek-chat   or   deepseek-reasoner
Remember key   off = session memory only; on = saved in local localStorage

Judging a “free tier”

Four things decide it. Unit of measure: by run (N images or N songs a day), by token or character (APIs and TTS), by credit (common in image and video generation, where different operations cost different amounts) or by duration (transcription). Reset cycle: a daily reset is the friendliest, a monthly one needs planning, and a one-off grant ends when it is used up. Quality cuts: free tiers often hand you low resolution, a watermark, an older model or a long queue — the free outputs of remove.bg and Clipdrop only reach preview resolution. Hidden conditions: a card on file (the OpenAI API free credit was discontinued and now needs a top-up), an invitation, an education email address. The only genuinely unlimited free option is running an open model locally, and the price is your hardware and your electricity. When choosing, look at a leaderboard for quality first, at Artificial Analysis for price, and then test the free tier on a real task.

Works in China, foreign phone numbers and regional limits

Three ideas that get mixed up constantly. Network reachability: whether the site opens over a direct connection from China — OpenAI, Anthropic and Google products are not reachable by default. Account sign-up: even when you can reach it, Claude asks for a phone number from a supported region at sign-up, ChatGPT used to ask and has since relaxed it, and Gemini depends on a Google account that is opened region by region. Service region: being able to log in does not mean the service is available, and some products refuse service by IP. The table only marks “Works in China” for products whose network, account and service are all unrestricted, and lists the sign-up barrier separately as “Foreign phone”. Chinese-made models (DeepSeek, Qwen, GLM, Kimi, Doubao, Hunyuan) have reached the top tier on most leaderboards, so everyday work does not have to insist on an overseas product.

BYOK and where your data goes

There are three places your data can go when you use an AI service. The provider's own product (the ChatGPT web app, the Kimi app): content is stored by the provider and is training data by default in most cases, which you can switch off in the settings. A third-party wrapper: content passes a third-party server before reaching the model vendor, adding one more leak risk, usually without telling you. BYOK tools (this site's AI Chat, Continue, Aider, Open WebUI pointed at a cloud API): your key, your request, sent from your device straight to the provider with nobody in between; the transcript stays on your machine. Keys on this site live in session memory by default and are written to localStorage only if you tick “Remember key”. Local running (Ollama, LM Studio, Whisper, ComfyUI, this site's cutout / OCR) keeps data entirely on the device, and is the only option without reservations for contracts, medical records or unreleased code.

What hardware local models need

Text models are estimated from parameter count: a 7–9B model at 4-bit quantisation takes about 5 GB of memory and runs on a laptop; 14B takes about 9 GB; 32B takes about 20 GB and needs a discrete GPU or a Mac with 32 GB or more of unified memory; 70B and above is workstation territory. Ollama and LM Studio pick a GGUF quantisation level automatically and split work between CPU and GPU. Image models: SDXL needs 6–8 GB of VRAM, FLUX is comfortable above 12 GB; video models generally want 16–24 GB. Whisper transcription runs in real time on the CPU through whisper.cpp. Unified memory makes Apple silicon very good value for local inference; on Windows, prefer an NVIDIA card, since the CUDA ecosystem is the most complete.

How to read a leaderboard

LMArena runs anonymous battles with user votes, which reflects “which answer people prefer” and is most useful for chat tasks — though it favours verbose, prettily formatted answers. LiveBench rotates its questions to fight training-set contamination; SWE-bench and the Aider leaderboard measure coding on real tasks; Artificial Analysis shows quality, speed and price on three axes and is the first place to look when picking an API provider; OpenCompass and SuperCLUE cover Chinese-language tasks. Any single leaderboard can be gamed, so pick a model from the overlap of two or three and then test it on your own real tasks.

Agents, workflows and MCP

Going from “chatting with a model” to “letting the model do things for me” takes three ingredients: tool calling (the model can trigger a search, read a file, call an API), multi-step planning (deciding the next move from the result) and state plus memory. Dify, Coze and Flowise assemble those pieces visually, n8n and Zapier grew AI nodes out of the automation-tool direction, and LangGraph, CrewAI, AutoGen and the OpenAI Agents SDK are the framework route. MCP (Model Context Protocol) solves a different problem — how tools plug in uniformly: a tool exposed by one MCP server works in Claude, Cursor, ChatGPT and any other client that speaks MCP, instead of being written once per client.

Inclusion criteria

reachable on 2026-09-09, an identifiable product, and something concrete to say within its category; 8–18 entries per group. Wrapper sites, pure marketing landing pages, and products that need an invitation code with no public sign-up are not accepted.

Labels

accurate as of the pricing page and sign-up flow on the day they were checked; AI products change every month, so the product's current page wins where they disagree.

Reviews are the editor's opinion

strengths and weaknesses come from leaderboards and hands-on use, and do not guarantee the same result on your task.

What is not included

vertical-industry AI SaaS (legal, medical, financial), AI hardware, platforms that require enterprise procurement, and single-model promo pages.

Boundaries of the on-site tools

BYOK tools need your own key and provider account; local models are limited by browser memory, and the cutout and OCR models download tens of MB on first use.

The directory data loads statically with the page; search, filtering and bookmarks happen locally in the browser, and this page sends no requests and records no clicks. Outbound links open in a new tab with rel="noopener noreferrer" and are then governed by the other site's privacy policy. The on-site AI tools are BYOK: the key and the conversation travel only between your browser and the provider you chose; the local-model tools (cutout, OCR, face blur) never let your data leave the browser.

• LMArena leaderboard（访问日期：2026-09-09） — model rankings and method notes from anonymous head-to-head voting.
• Artificial Analysis（访问日期：2026-09-09） — quality, speed and price on three axes; the place to start when picking an API provider.
• Ollama README（访问日期：2026-09-09） — installing open models locally and the memory they need.
• OpenAI API reference（访问日期：2026-09-09） — the request format that “OpenAI-compatible” endpoints follow.
• Model Context Protocol specification（访问日期：2026-09-09） — the open protocol for connecting tools.
• Hugging Face Open LLM Leaderboard（访问日期：2026-09-09） — standardised evaluation for open models.