Getting started
Open a workspace, configure your LLM provider, and send your first chat in Ask, Plan, or Agent mode.
Quick topics
The features and agents pages link here for deeper reading.
Workspace and editor
The main window is IDE-style: a left navigation rail (workspace, projects, agents, tools and logs), a center area with Chat, Editor, and Markdown preview pages, optional guidelines and utility panels, and an integrated PTY terminal at the workspace root. Panels can be hidden, resized with splitters, and detached into floating windows; layout is saved per workspace.
Roadmap entries in the nav (multi-agent workspace, execution plans, knowledge base, The Brain) are planned placeholders with no backend—see the roadmap.
Ask, Plan, and Agent modes
Ask is read-oriented: inspect files, search, and git status without writes or shell. Plan can save plan markdown under .plan/ but not arbitrary writes or shell. Agent adds patches, writes, and shell for implementation work. The mode applies to the next run and controls which tools are advertised to the model.
By default, writes and shell working directories stay inside the open workspace folder. Each session can opt in to modifications outside the workspace—application-level policy, not an OS sandbox.
Chat, sessions, and models
Each workspace supports multiple parallel chat sessions with separate conversation history and agent engines; switching tabs does not stop a run in another session. Configure provider, API key, and models in Settings; attach up to four images per message when using a vision-capable model. Debugging mode controls full tool JSON in chat; Tools & Logs and Execution Detail hold the full trace.
Agent loop
After you send a message, a background thread runs a ReAct-style loop: build system prompt (mode, workspace, tools, project rules), call the LLM, run permitted tools locally, append results, and repeat until a final reply, cancel, error, or step limit. Unexpected tool failures return error JSON so the loop can continue.
Read how the agent works → · Architecture summary on the site →
Project rules
Place Markdown files in <workspace>/.rules/*.md. The app loads all of them into the system prompt for Ask, Plan, and Agent. Edit via Workspace → Project Rules…. Rules add project-specific guidance; they do not replace mode gates or the write sandbox.
Related pages
- Download and build from source
- Feature overview
- Agent architecture
- Planned work
- Product feature inventory (maintainer evidence)