Sona

Streaming chat. Edit a question, regenerate a reply, or branch the thread. A reply can show which tool ran.

Custom assistants. Save instructions and an optional preferred model.

Sona Auto. Chooses among reachable local models and connected cloud models, and shows why.

Paid models. Off unless you choose ask-first or allow, and a monthly budget can stop them. OpenRouter chat was verified. Live replies on Anthropic, Gemini, Groq, and DeepSeek with a working key were not.

ChatGPT plan. Plan-usage sign-in was verified on this private app. Agent mode is not offered on that connection, and eligibility can change.

Image questions. Images go only to models that accept them. Text-only models refuse the image.

Voice in chat. Local transcription and spoken replies, with live captions, barge-in, and mute. Turns stay in the chat.

Still images. Generation, editing, inpainting, and upscaling run when an image worker is connected. Still images were verified in production.

Video. Text-to-video and image-to-video stay off. The acceptance check skips them until a real clip renders.

Web research. Reads public pages and cites them. Private and loopback addresses are refused.

Document questions. Answers from uploaded files cite the passage they used.

Reviewed plugins. Calculator, clock and dates, and unit conversion install from the catalogue and run as slash commands or agent tools.

Tool choice. Stay on Auto, or pin chat, web search, create image, or edit image.

Consented sessions. Browser, desktop, or a virtual computer starts only after you set a scope and a time limit. You can pause, take over, or stop.

Approvals. The model cannot fill credentials or payment fields. Submit-like actions wait for you.

Sona Computer. Screenshots, clicks, typing, keys, scrolling, and opening sites on a virtual computer you attach. Checked on the real machine and through the production app on 31 October 2026.

Choosing a click. A model needs both vision and tools to pick a click from the screen. Text-only models can still open sites, type, and press keys.

Busy virtual machines. A stall is reported as a timeout, with status and recovery. Restore and force power-off wait for your confirmation.

Projects. Each project keeps its own instructions, files, and chats.

Project Brain. Stores facts you have verified for that project.

Memory scopes. Global, one assistant, or one project. An assistant can keep new memories to itself.

Files. Text, Markdown, code, CSV, JSON, HTML, PDF, Word, and images, 25 MB each by default.

Organizations. Invited members can share projects, assistants, and chats, with roles, budgets, an audit log, backup, and a hand-off when someone leaves.

Agents. A run records its steps. Coding tasks can use Claude Code with your own Claude plan; that path was verified.

Workflows. Each step is saved, and a run resumes after a restart.

Steps. Branch, repeat within a fixed cap, and call plugins, approved MCP tools, or your own webhooks.

Approvals inbox. A step that needs permission waits. A denial stops the run.

Desktop triggers. Folder watches and inbound triggers run in the desktop app. The hosted site does not expose them.

Local Mode. On the desktop app, chat, files, and history stay on the device, and cloud providers stay disconnected.

Local models. Sona discovers local runtimes and the models they offer.

Your computers. Pair your own machines, place a job by what each one can do, and continue it from a checkpoint if one drops.

Where computer use runs. Virtual-computer and desktop actions run on a machine you pair. They do not run in the website.

Account separation. One account cannot read another’s chats, files, or projects. Exports leave secrets out.

Sandboxed tools. Plugins and MCP tools run in a sandbox. The agent only sees tools you approved.

AnxLabs connectors. AnxOS status reads, and start, stop, or restart after approval. AnxGraph can supply labelled context. An approved proposal can go to Agent Office.

Shared credentials. A credentials service was reviewed and not connected. Sona does not use it.