Agent flows with real tools
Flows are first-class. Agents can call dozens of tools across MCP, web search, web scrape, spreadsheet/CSV intelligence, sandbox execution, and graph-oriented tooling you can wire into the flow itself — including visualization/line-graph style outputs when the job is to show the shape of the data, not just narrate it.
Builders: flows and front ends
- AI flow builder — compose agent graphs instead of one-shot prompts.
- Figma Make–style UI builder — generate front ends when the deliverable is interactive, not a memo.
- Multi-asset creation in chat — drafts, tables, scripts, and UI pieces from one session.
Chat is the center; editors are the side doors
The layout is intentional: chat in the center, file/folder browsing on one side, code and editor surfaces on the other when you need them. There is a code editor and script runner, plus a sandbox shell for power users. It is not trying to win a VS Code bake-off. It is trying to help analysts and makers get to a verified output fast.
Corpus storage and workspace mounts
- Org-scoped uploads — sticky corpus files into on-device storage (native data directory or local Docker volumes), with graph database + vector database backing for documents and CSVs.
- Workspace mounts — load a folder from your computer into the sandbox environment and work against that project.
- Configurable sandbox — choose which Python/NPM packages are available for analysis and builds.
Research without browser clutter
A no-frills in-app browser (no history, no bookmarks) keeps research in the same thought-stream as the agent session. Pure flow: look something up, bring the signal back into chat, keep moving.
What leaves the machine
Everything stays on your desk. Org corpus lives in the local data plane — OS-sandboxed processes in the native build, or Docker volumes in the Docker build. The primary egress is model context for the LLM calls you enable — not “your whole hard drive uploaded to a consumer chatbot.”