A prompt goes in. A file comes out. Everything in between is visible.
QGI Studio is built around three nouns. A project is a workspace with real files. A run is one turn of the agent inside that project, with an engine, a model, a mode, and a narrowed toolbox. An artifact is what the run leaves behind: a deck, a page, a sheet, a document, a video, a workflow.
One bar, four decisions.
Every project starts in the same bar you see on the home screen. Before you press Send you choose:
- Template. None, or one of 34 starting points (a Credit Memo, an Operating Review Deck, a Monday Morning Brief).
- Session mode. Chat for a quick answer, Plan to scope before building, Agent for a complete autonomous run that writes files and runs tools, FactResearch for sourced answers with citations.
- Engine. The coding agent that will execute:
claude-code,cursor,cursor-cli,Quark,Codex,opencode,kimi,QGI. - Model. The model catalog for that engine, refreshed live from the fleet. On
claude-code:Fable 5,Opus 5,Sonnet 5,Haiku 4.5.
Or skip the bar. Pick a file type from the rail (Slide deck, Document, Web page, Sheet, Image, Video, Agentic Workflow, Sketch, Workflow, Data (JSON), Script) and the project is scaffolded for it.
Watch it think.
The left panel is the agent's working log, not a chat. Each turn shows the engine that ran it, elapsed time, and a collapsible trail of Thought, tools, result: which tools were called (read, shell, grep, getMcpTools, mcp, edit), whether each succeeded, and what the agent decided next. When it finishes, the files it produced are listed with Open buttons.
It is honest about failure. If a conversion tool can't decode a format, you'll read that in the log and see the workaround it chose.
Preview, code, edit. And check it on a phone.
The right panel is a tabbed workspace. Open any artifact in Preview, flip to Code to read the source, or Edit to change it by hand. The preview toolbar gives you desktop, tablet and mobile widths, zoom, reload, open-in-browser, full screen and export.
Dashboard, mobile preview, transcript, versions, tooling.
Every run is a version.
Project history records each run with its prompt and run id; save a named version whenever you want a checkpoint, and roll back when you don't like where a run went. Comments sit beside the artifact. Share hands the project to a teammate; Save as forks it.
A smaller toolbox means sharper choices.
Each project can follow the studio's full toolbox or choose its own. In Finance Studio that's 62 collection tools, 6 MCP servers and 10 skills. Toggle them per project so the agent reaches for filos_decision_whatif instead of guessing. Skills and plugins can be added for a single project without touching the rest.
Turn a run into a schedule.
Anything you've done once can become a flow file: a prompt on a cadence, a governed flow with approval gates, or a deterministic HTTP job. Project workflows deploy straight from their project; governed ones ship through review.