Assistant

The Assistant (in the sidebar, or ⌘ J; the sparkles on a phone) opens a chat on the right of the dashboard. A new chat offers a few suggestions to start with one click, or ask in your own words:

• "Write a LinkedIn post about our new pricing and schedule it for Monday 9:00"
• "What is scheduled this week?" or "Which posts failed, and why?"
• "Create a collection for Instagram that posts every weekday at 9:00"
• "Move all drafts tagged launch to the Launch collection"

It works with the same tools AI agents get over MCP, as you and with your role: a viewer's assistant can only look. Looking things up happens at once. Every change waits for you: a card says what it is about to do (the post's text, the time, the channels) with Allow and Deny. Posts it writes are drafts unless you ask for a time or to publish now.

Add pictures, videos, sounds or text files to a message with the paperclip, by dropping them on the box, or paste a screenshot. Pictures, videos and sounds go to your media library (so the assistant can put them on a post) and the model sees the pictures; a text file (.txt, .md, .csv, .json...) goes in the message as its text, up to 100,000 characters. The assistant's model has to read pictures for it to see a screenshot.

Your chats are kept on the server, yours alone in each workspace: Chats (the clock) lists them newest first, searched by title, and opens one where you left it, a change still waiting for Allow included; New chat starts another. The panel's left edge drags to make it wider, and Full screen gives the chat the whole window (Esc goes back). What you type and what the tools find (post texts, names) go to the AI service to answer; secrets don't (it has no tools that show API keys, webhook secrets or channel tokens, and none that connects a channel or deletes the workspace).

Where it runsCopy Anchor Link

On the server's AI provider, like the design editor's AI tools: staqly cloud includes it. On your own server, either:

• AI_FAL_KEY: fal.ai's router, where AI_ASSISTANT_MODEL picks any OpenRouter model (Gemini by default, or anthropic/claude-sonnet-4.5, openai/gpt-5...), or
• AI_ASSISTANT_BASE_URL (+ AI_ASSISTANT_API_KEY, AI_ASSISTANT_MODEL): any OpenAI-compatible API, such as OpenAI, OpenRouter, DigitalOcean's serverless inference, a local Ollama or your own LiteLLM.

See configuration. Without either the Assistant is not shown. Each answer counts as AI usage of the workspace: what the endpoint reports it cost, or else its tokens at the model's price (AI_ASSISTANT_PRICES for that endpoint, else the built-in price list, which AI_PRICES_FILE extends).

When it can't answerCopy Anchor Link

AI_ASSISTANT_FALLBACK_* sets a second, usually cheaper model. It answers when the workspace's monthly AI allowance has run out, and when the first model refuses before answering (its own quota ran out, it is down). A model that already started answering is not replaced halfway.

When neither can answer and the server names a model for the device (AI_ASSISTANT_BROWSER_MODEL), the chat offers Continue on this device: the assistant then runs in your browser (it needs WebGPU: Chrome, Edge, Safari 26 or later). It is free and nothing leaves the browser, but it only writes text: it can't read or change anything in the workspace. The first time downloads the model (a few GB); New chat goes back to the server.

Pictures and videosCopy Anchor Link

Ask for a picture or a video ("make a 9:16 video of a sunrise over the city") and the assistant makes it with the server's AI tools (generate_media), after you allow it. A picture takes about a minute, a video a few; it then offers to keep the result in the media library, ready for a post.