Projects - agentic coding in chat

Open a project folder and the conversation becomes a workbench: your AI reads files, proposes edits as colored diffs, runs builds and tests with output streaming live, and asks before every action. Everything it does lands in the same signed conversation record as everything it says.

Setup: the Your Own AI Build add-on (once)

Projects are powered by Your Own AI Build - an optional, free ~50 MB add-on that isn't part of the base install. Download it once and projects light up everywhere; it also runs tools in chat. Two places to get it:

• The projects dropdown in the header - open the folder menu and the download offer is right there. It installs in the background while you keep chatting.
• Add-ons › Components - the Projects card. Settings › Storage is where you see what is installed and remove it.

Your Own AI Build is our open-source fork of xAI's Grok Build, rebuilt local-first: it talks only to your machine, with upload, telemetry, and auto-update removed. Source at WeAreFlowsta/Your-Own-AI-Build (upstream: xai-org/grok-build), Apache 2.0 licensed - every claim verifiable in the code.

Opening a project

• The folder button in the header,
• "Open a project..." in the sparkle menu, or
• drop a folder onto the message field.

The header shows the project chip with a status dot while it's open.

Permissions, not surprises

By default every command and edit asks first. One choice - Ask, Auto, or Approve everything - decides how much a project may do unasked, per project and switchable live; whatever you choose, every action lands in your records. See Permissions and approvals.

Checks after edits

When the agent thinks a turn is done, the project's own check command runs first - a typecheck, check or lint script from package.json, cargo check, go vet, or ruff - and failures go straight back to the agent to fix before it finishes, two rounds at most. The outcome is written on the turn: "Checks passed (npm run typecheck)" or the errors that remain. Projects without a check command are left alone.

Undo this turn

A finished turn that changed files offers Undo this turn's changes under its steps: every edit is put back, files the turn created are removed, files it deleted are restored - in any folder, with or without git. What was undone is written on the turn and kept in your records. This is what makes "Approve everything" safe to pick.

Watching the work

Each working turn shows its whole story live and folds to one line when it ends - and you always know what it's doing while it runs. See The working session. A "Simple project view" setting trims the verbosity, never the liveness.

The project remembers

Decisions and conventions save to project notes all your AIs share - the AI can ask to save a note (you approve each one), and each session's takeaways are distilled when you finish. Reopening a project conversation picks up where it left off.

Proving what happened

Export the conversation - steps, diffs, and commands included - as a file, optionally signed so anyone can verify it. See Export and receipts.

Good to know

• Project work only routes to models that can drive tools; the app offers a switch rather than silently substituting.
• An agent that opens a picture or a PDF with a model that cannot see does not lose the turn: in an automatic mode the call goes to a downloaded model that can see, when one fits; otherwise the image becomes a note the AI can act on (a PDF can be read as text), and the rail says why the picture was not looked at.
• Two cost levers in Settings → Routing keep project side-work - or whole sessions - on your device when your hardware fits them.
• Shell commands an AI writes can hand off to your own terminal, pre-filled at your prompt - you stay the one who presses enter.