Why do the inline buttons take a long time to load?
If you have not added an API key, some actions may try to use on-device processing. That can be slow depending on your computer, browser, memory, CPU, GPU, and driver support. This kind of local browser AI is still experimental, so the first request can feel heavy on older hardware.
For faster and higher quality results, add an API key from OpenAI, Google AI Studio, DeepSeek, OpenRouter, or another supported provider in the sidebar settings. Ceres rewrites, speech transcription, and browser assist become much faster and more reliable with a good provider key.
Why does it stay on connecting or processing?
Chrome extensions run under Manifest V3, and Chrome may suspend extension workers to save resources. If the extension says connecting or processing for a long time, it may be waking up before it can handle the request.
Give it a moment, especially after the browser has been idle. After the worker wakes up, the next requests are usually faster. If the same problem continues, send us a bug report with your browser, extension name, model provider, and what you were trying to do.
How should I set up my API key?
If you use OpenAI, GPT-5 mini is a good everyday choice for writing and assistant work. For Speech to Text in the browser, choose a speech model such as GPT Audio when it is available in your provider settings.
If you use Google AI Studio, Gemini 3 Flash is usually enough for inline proofread, short rewrites, transcription cleanup, and normal browser tasks. Gemini 3 Pro can be useful for heavier Ceres work, but it is often overkill for simple proofreading or short replies.
For lower cost, try DeepSeek, OpenRouter, or another OpenAI-compatible provider. Some providers include free usage, budget models, or very low pricing. If you want the tool to feel instant, choose a fast model before choosing the biggest model.
What should I use with Ceres inside an IDE?
Ceres works well with local AI through Ollama and LM Studio, and it also works with OpenAI-compatible APIs. If you want to save money, a previous generation coding model or a budget reasoning model can be a smart choice.
Ceres is role/chat based. It is not built around long agentic sessions that keep calling tools in the background. That means models that still write clean code directly can work very well here, especially for selected files, focused edits, debugger context, and controlled IDE help.
If you want the safest path, use a reliable model from OpenAI, Anthropic, Google, DeepSeek, or another provider you trust. If you want local control, test Ollama or LM Studio with a coding model that fits your machine.
Drag and drop is not working on Visual Studio Code
When you drag editor tabs inside Visual Studio Code, VS Code may block the normal drop behavior. Hold Shift while dragging the editor tab into Ceres. This lets the file drop into the sidebar correctly.
How does image generation work in Ceres?
Image generation runs during the conversation. If you have an image project open, generated images can also be saved there. You can create a whole set of images and keep working without manually saving every file.
Image generation currently works through the OpenRouter provider. Choose the image model you prefer from the supported list, including models such as Nano Banana, Flux Dev, or another OpenRouter image model when available.
What is the context window bar?
After you add a key, you may see a bar or window at the top that slowly fills up. This is the context window. It shows how full the conversation is getting as you add text, files, URLs, PDFs, images, or long replies.
If the context window becomes too full, the model may stop seeing the whole conversation clearly. Use the arrows and Break Window to split the conversation at the point you choose. After that, the model focuses on the lower, useful part of the chat. This lets you keep working without deleting the whole conversation or paying for very large context models every time.
Why can some large files or URLs not be read?
Ceres includes read limits to protect you from accidentally sending huge files and creating expensive API calls. If a file or URL is too large, the extension may stop before reading all of it.
Open settings and increase the Read URL / Read File limit if you know the file is safe to process. Use a larger limit only when you understand the possible token cost.
Can I send images, PDFs, and other files?
Yes. You can send images when your selected model supports vision. You can also send PDF files and read PDFs that are already open in the browser or inside the IDE workflow.
For image understanding, choose a multimodal model. For PDF reading, keep an eye on file size and context usage, especially with long documents.
Can I edit images with AI?
Yes, when the selected provider and model support image editing or image generation workflows. You can ask the AI model to transform an image, follow instructions, and create a new image based on one, two, or three previous images.
For direct cropping, resizing, conversion, and local background removal, Luna Crop is the better tool. For generative image changes, use Ceres with a provider and model that supports the image task you want.
Why do I need my own provider key?
Most AI extensions do not own the large AI models they use. They call providers such as OpenAI, Google, Anthropic, DeepSeek, or other model platforms, then include that cost inside a subscription.
We chose a different approach. You bring the provider you want, use the model you prefer, and pay the provider directly for the usage you create. This avoids hiding model costs inside our pricing and lets you use free provider allowance, local models, or budget APIs when they fit your work.
How much will API usage cost?
It depends on the model and the product. Ceres usually makes many small calls for proofreading, rewriting, transcription, or short browser text actions. It does not send an entire long session for a simple correction. With a mini or budget model, the monthly cost can be tiny, sometimes under one euro for light use, or free when your provider includes free allowance.
Ceres can cost more because it may analyze longer conversations, files, URLs, PDFs, code, or project context. Even then, it can still be much cheaper than long autonomous coding sessions that call the model repeatedly while exploring the project.
A practical starting point is to add five or ten euros of credit to your favorite provider and watch usage for a few days. If you choose Gemini Flash, DeepSeek, or another budget model, normal use may feel almost invisible in cost.
Why is Speech to Text giving the wrong text?
In the browser, Speech to Text works with the browser speech-to-text engine unless the user has configured an audio LLM in settings. The browser handles language detection automatically.
Browser Speech to Text
If the result is completely unrelated, or you receive recording errors, start with the microphone. Check browser permission, selected input device, microphone volume, and the physical connection. Test with a simple audio recorder until the recording sounds clean.
VS Code and IDE Speech to Text
Inside the IDE, Speech to Text can fall back to your operating system speech recognizer when the remote audio API is not available. This is useful, but it depends on the speech packages installed on your computer.
If you see a message like poorly transcribed on VS Code, configure audio model at API settings, the local OS recognizer probably heard something but produced low-quality text.
If the output is in the wrong language, your operating system may be using a different recognizer than the language selected in Ceres. On Windows, open language and speech settings, add the language you want, and install its speech recognition package. On macOS, open keyboard speech input or Speech settings and make sure the language you want is available for speech recognition. After installing the correct package, restart the IDE and try Speech to Text again.
If you are on Linux and see OS speech recognition is not available on Linux. Please use an audio model at app settings., this is expected. The local OS speech fallback is available for Windows and macOS only.
If transcription quality is still not good enough, you can optionally connect your own audio LLM in the application settings. When an audio model provider is configured, audio processing can use the LLM instead of relying only on the browser or operating system recognizer.
Why do I see "On-device AI is not available on your computer"?
Ceres tries on-device AI when no AI provider is set in settings. This lets some AI features run through Chrome on your own computer.
Your computer must support Chrome built-in AI. Chrome docs list two supported paths: a suitable GPU, or a CPU setup with at least 16 GB of RAM and 4 CPU cores. The first run may also need time to download the local model. Source: Chrome built-in AI requirements.
If Chrome cannot run on-device AI on your computer, add an AI provider in settings. After that, Ceres will use your selected provider for these AI features.
Which features use on-device AI?
On-device AI is not used for every Ceres feature. It is mainly used for lightweight rewriting tasks, such as grammar and spelling fixes, shorter text, longer text, summaries, and custom rewriting prompts when your computer supports it. It can also be used for audio transcription and OCR.
Sidebar chat, Live Assistant, and Command Assistant do not use on-device AI. Command Assistant is the voice feature that rewrites selected text from your spoken instruction. These features need an AI provider because they handle more demanding context, voice, and assistant workflows.
What engine is behind on-device AI?
Ceres uses the browser built-in Prompt API when on-device AI is available. In Google Chrome, this API runs on Gemini Nano. In Microsoft Edge, Microsoft uses Phi-4-mini instead. Sources: Chrome Prompt API and Microsoft Edge Prompt API.
The browser provides the model and decides if your computer can run it.
Can I use the products for free?
Yes. You can use our products for free for personal work and professional work. Free usage includes a limit, but the limit is designed to be useful for normal daily tasks.
If you find yourself using the tools all the time, then it really makes sense to purchase a license.
Where can I buy a license?
We offer a license key through Gumroad at a fair price. After purchase, you can use the product without the free-use limits. You may use your key on all your own devices. If an extension supports both the browser and IDE, the same key unlocks both.
Do not publish or share the key with strangers. If a key appears on the internet, we may need to disable it. Sharing with your close family circle is fine. Thank you for supporting the work.
Why does overlay not work when I drop an image into the Luna Crop sidebar?
If dropped images are not staying as overlays, open Luna Crop settings from the gear icon and enable overlay images.
When overlay images are disabled, the default action is to clear the temporary image after you crop or drop a new image into the sidebar.
Why is Luna Crop background removal slow or crashing?
Luna Crop background removal uses on-device AI. On older computers it may fall back to CPU processing, which can be very slow. On some older systems, the processor or browser may not support the required features, which can lead to messages such as worker crash, SIMD not supported, or similar errors.
Try updating your browser, graphics drivers, and processor drivers. If possible, test on a newer machine or a computer with better GPU support. At the moment, we do not provide cloud background removal for Luna Crop unless we build a dedicated cloud service later.
Why is Luna Crop upscale very slow?
Luna Crop upscale uses on-device AI. To increase resolution, the AI must infer likely extra image detail and rebuild a larger version of the image, so the process can be especially slow.
Upscaling starts only after you select the specific area you want to make higher resolution. We use this method so the AI has less image data to process and can finish faster. If needed, you can still select the whole image and upscale everything.
If you do not have a strong graphics card, avoid relying on upscale in a production workflow. Our app does not use a cloud provider for upscale or background removal, so both tasks depend on your local machine.