AI Tools Directory

The product that brought AI chat to a general audience, and still the first one most people try. Its strength is breadth: plugins, custom GPTs, a code interpreter and web browsing all live behind one interface, so a task needing several capabilities does not send you elsewhere.

The weaknesses are equally clear. The gap between the free and paid models is substantial, and on long documents or tasks with strict formatting requirements the output tends to drift from the instructions partway through. Its Chinese is competent but reads less naturally than the domestic models — worth comparing directly if you write Chinese copy.

If you are in mainland China: neither the site nor the app is reachable, and signup requires an overseas phone number. Most of the “direct access” mirrors you will find are relays that route your conversation through a third-party server first — worth knowing before pasting anything sensitive.

The model to reach for when a task requires reading a lot before acting: a fifty-page document, an entire codebase as context. It holds the structural relationships better rather than just responding to your last sentence. Its prose is also restrained — less of the instantly recognisable AI cadence.

Code is its strongest area, particularly edits inside an existing codebase: following the conventions already there instead of rewriting things its own way.

The limitation is breadth. No image generation, weaker web search than ChatGPT or Gemini, and essentially no plugin ecosystem. It is a very capable single model rather than a platform.

If you are in mainland China: the region is not supported and the site redirects to an unavailable notice.

The context window is its most concrete advantage — a whole book or several hours of meeting notes fit in one go, with no chunking and stitching. Its multimodal handling is also the most complete of the major assistants: images, audio and video all work as direct input.

Integration with Google’s own products is the other reason to pick it. If your material already lives in Gmail, Drive or Docs, it can pull from there rather than making you export and paste.

The trade-off is consistency. Run the same prompt a few times and the variance is more noticeable than with ChatGPT or Claude, especially on tasks with strict formatting requirements. It suits “help me work through this pile of material” better than “polish this paragraph until it ships”.

Not reachable from mainland China.

The most consequential Chinese model of the past two years. It reaches close to frontier-level reasoning while charging a small fraction of what comparable APIs cost — for individual developers and small teams, that difference often decides whether an idea is viable at all.

The weights are open, so it can be self-hosted. Where data cannot leave your own network, that is not a nice-to-have but the deciding factor. Its Chinese is natural, and it holds context across long Chinese documents more reliably than most international models.

The gap is in ecosystem: no plugin system, weaker multimodal capability than GPT or Gemini, and a fairly plain web interface. It suits “give me a correct answer” or “write this function” far better than workflows that need to orchestrate external tools.

Reachable from mainland China without any additional setup.

The Chinese assistant that has invested most in long-document handling. Upload a few hundred thousand characters and ask questions, and it locates specific passages and connects information across chapters more reliably than most alternatives at this price. Well suited to research summaries and working through reference material.

Its Chinese reads naturally, usually without needing the word order fixed afterwards. Web search works inside China, which makes it more practical for Chinese-language sources than international models that cannot reach domestic sites at all.

The gap is in reasoning depth: multi-step derivations and problems needing a strict logical chain still fall short of the frontier models. Code is not its strength either.

Reachable from mainland China without any additional setup.

The first to reach wide adoption and still the one with the broadest editor support: official plugins for VS Code, the JetBrains suite, Neovim and Visual Studio. When a team is split across editors, it is often the only option everyone can actually use.

It is at its best continuing what you started — you write the signature or a comment, it fills in the rest. Repetitive code (test cases, data transforms, boilerplate config) is where the time saving is most obvious.

Worth knowing going in: the completions look plausible without always being correct, particularly around edge cases and error handling. Treat it as a fast but careless pair — every line still needs reading before it stays.

Reachable from mainland China, though speed depends on your connectivity to GitHub.

A VS Code fork that turns AI from a completion plugin into a way of working. The key difference from completion tools is that it reads the whole project: you can say “make this function’s error handling match the rest of the codebase” and it will go look at the rest of the codebase.

Multi-file changes are where it earns its keep — renaming a concept, adjusting an interface and updating every call site. That work is far faster here than pasting back and forth into a chat window.

The cost is switching editors. If your settings, keybindings and extensions live elsewhere, the migration is real, and the VS Code version it forks from sometimes lags upstream.

Reachable from mainland China, though signing in occasionally takes patience.

Different from both completion tools and AI editors: it runs in your terminal and decides for itself which files to read, which commands to run and what to change, then shows you the result. You describe the task; it does the work, rather than handing you a snippet to paste.

It fits tasks that span several files and require understanding the current state first: tracking down a bug, adding tests to existing behaviour, migrating a batch of code to a new convention. It reads how things are done before changing them.

Worth being deliberate about: it can execute commands. It asks before each step by default, but understand what it is capable of before loosening that — especially in a directory with uncommitted work or production configuration.

It shares Claude’s account system, so mainland China is likewise unsupported.

Judged purely on how the images look, it is still the most reliable: lighting, composition and material all land well from an ordinary prompt. For concept art, posters and mood references it saves the most time in sifting through results.

The cost is control. Precise instructions — this object in the lower left, that figure facing right — often produce a more beautiful image that is not the one you asked for. For work that must match a design exactly, the Stable Diffusion route with stackable conditioning fits better.

It is also closed and cannot be self-hosted, and access is subscription-only with no lasting free tier. Worth knowing before you start.

Not reachable from mainland China directly.

Being open is the defining difference: the model runs on your own machine, images never touch a third party, and there are no content policies imposed on generation. For commercial work, or when the source material is confidential, that alone often settles the choice.

The ecosystem is the other advantage. ControlNet-style conditioning constrains composition by pose, depth or edge map; LoRAs train a specific style or character from a handful of examples. “The same character in every image” or “match this line art exactly” is achievable here and difficult in Midjourney.

The cost is setup. You configure the environment, pick models, tune parameters — and running locally needs a capable GPU. Out of the box it also looks worse than Midjourney; good results depend on knowing what to adjust.

The site opens from mainland China.

The most complete option among those usable inside China without setup: image generation, image editing and video generation all in one product, with no shuttling between tools.

The practical advantage for Chinese users is that prompts need no translation. International models understand Chinese prompts less well than English ones, while this is trained on Chinese material and handles culturally specific descriptions without drifting. Chinese typography also renders far better — international models tend to produce character shapes that are almost, but not quite, real characters.

The gap is overall image quality against Midjourney, and complex scenes are more prone to structural errors.

Reachable from mainland China; signup requires a phone number.

Not just text-to-video but a set of post-production tools: rotoscoping, object removal, motion tracking, turning a still into a shot with camera movement — all in one interface. For people already producing video, it replaces several applications that would otherwise each need opening.

The generation side is also more controllable than most: camera movement can be specified, and an image can anchor the first or last frame to constrain where the shot goes.

The limits are duration and consistency. Individual clips are short, and stitching several into something with a narrative struggles to keep characters and settings stable — a problem common to every current video model, not specific to this one.

The site opens from mainland China, but signing in and generating depend on an overseas connection.

Among the video generators usable directly from within China, this is one of the stronger ones. Human motion holds up reasonably well — hands and walking gaits, the places where these models most often break down, fail less often here than in most alternatives.

Both text-to-video and image-to-video are supported. Animating a supplied still is easier to steer than describing a shot in words alone, and remains the more controllable approach across every current video model.

The weaknesses match the category: short clips, poor consistency between them, and structural breakdown under complex camera movement. Chinese prompt understanding is a genuine advantage for locally specific scenes.

Reachable from mainland China.

The most natural synthetic speech available right now: pauses, intonation and emotion come close to a real reading rather than the immediately recognisable flatness of most TTS. For audiobooks, video narration and podcast drafts, the output is often usable as-is.

Voice cloning is supported — a few minutes of material reproduces a voice. That capability is powerful and carries responsibility with it: using someone else’s voice requires their consent, and many jurisdictions regulate this explicitly.

Chinese output is less natural than English, occasionally carrying a non-native intonation. For Chinese-only material, domestic alternatives are worth comparing.

The site opens from mainland China.

Describe a style in a sentence and it returns a complete song: arrangement, vocals and lyrics. This was unimaginable a few years ago. For background music, short-video scoring or demo sketches, it produces something usable within minutes.

Lyrics can be written by you or generated. Chinese pronunciation has improved considerably from early versions but still has odd moments — generating a few takes and picking one is normal.

Licensing deserves attention: work produced on the free tier usually does not carry commercial rights. Check the current terms before using anything in a real project — these terms change relatively often.

Not reachable from mainland China directly.

It addresses the biggest problem with general chat models: you cannot tell whether an answer is real or invented. Here each claim carries a source link you can open and check. For information that leaves your desk — a report, a piece of research, technical reference — that difference is decisive.

Against traditional search, its value is reading the top results for you before answering, rather than handing you ten links. On a specific factual question the time saved is substantial.

It is not the tool for creative or long-form output; that is not what it is for. And a caution: having a source is not the same as having a good one. It cites low-quality pages often enough that the links still deserve a glance.

Not reachable from mainland China directly.