Everything below runs on your machine, against your local Ollama server. No feature phones home, because there's no home to phone.
The API dropdown (top center) controls how the app talks to the model. Both are useful; they solve different problems.
Each request stands alone; the model "forgets" the previous question immediately. This is the mode for tasks, not conversations — and it's the only mode that accepts images and audio. You can optionally carry the context tokens into the next request, which gives you conversation-like continuity even here.
The whole history is kept and sent with every message. The superpower: you can switch the LLM mid-conversation. Brainstorm with a fast 8B model, then hand the thread to a heavyweight for the final answer — the new model inherits the full context.
Chat mode can't receive images directly, so the app gives you a workflow instead of a wall:
The bottom of the app is built for real work, not one-liners:
Enter or the ▶ button sends everything, attachments included. Shift+Enter inserts a line break for structuring your prompt.Upload .jpg or .png files in Generate mode and analyze them with a vision model like llava. Diagrams, error dialogs, whiteboard photos, UI mockups — described, transcribed, or interrogated, without the image ever leaving your disk.
The built-in screenshot button shortens the loop: capture your screen and pipe it straight into the vision model. "What's wrong with this error message?" becomes a two-click question.
Upload .wav or .mp3 files — or hit the record button and speak — and analyze the audio with an audio-capable model like gemma4. Voice memos, meeting recordings, dictation: processed on your hardware, useful even when the content is exactly the kind of thing you'd never upload to a cloud transcription service.
Import .pdf, .txt, and .json files directly into the prompt context. Good for anything that fits in the model's context window: a config file, an article, a spec.
For large PDFs and text files, the RAG Tool builds local knowledge extraction over the document, then answers your questions from the relevant parts. A 200-page contract, a thesis, a technical manual — interrogated on your workstation, indexed on your disk, exposed to no one.
Flip on JSON mode and the model is forced to respond in structured JSON — not "usually JSON, occasionally an apology," but parseable output. This is the feature for people who pipe AI into other things: extraction pipelines, automations, form-fillers, anything downstream of the answer.
Let the model call local Python through tool/function calls. The model decides it needs a computation or a lookup, invokes the tool, and works the result into its answer — the agent loop, running entirely on your box.
When a response contains a Python, PowerShell, or Batch script, you can execute it straight from the app: pick the code block, review it, hit Run. The write-copy-paste-debug shuffle collapses into a click.
This is the feature no cloud chatbot will ever be allowed to ship. The SSH Bot turns Cryptic Thinking into an autonomous system administrator: it connects to a server over SSH with your credentials, executes commands, reads the terminal output, and reacts to what it sees on its own — chaining steps until the job is done.
df, sees the numbers, drills into the offending directory, reports back.Generate images locally with x/z-image-turbo and x/flux2-klein. Concept art, placeholders, illustrations — produced by your GPU, with no content policy server deciding what you're allowed to draw and no gallery of your prompts on someone else's servers.
Cloud UIs hide the dials. Cryptic Thinking hands them to you: temperature, context length, system prompts, and the rest of the model parameters — adjustable per session, with presets for the setups you keep coming back to. The model info view shows what the selected model supports (the capability chips turn green for vision, tools, and friends), so you always know what you're working with.