The Dune agent lives in the workbench, not in a browser tab. It reads your repo, edits files, runs the shell, and hands work to subagents, all inside permission rules you set. Every edit is checkpointed, so nothing it does is permanent until you keep it.
Pick how much the agent may do. Ask mode answers questions and never touches files. Agent mode edits and runs tools. Plan mode writes a checklist first and is read-only until you approve it. Debug mode hunts a bug down from the evidence. Projects can add their own modes as Markdown files.
- Plan mode can't write files or run a shell, whatever the permission rules say
- Custom modes live in .dune/modes/*.md, created with /create-mode
- One Hypery sign-in covers chat, inline completion and cloud workspaces
Big tasks split up. The agent hands parts of the job to subagents that run in parallel, each with its own role, its own context and its own permission rules. The chat shows them in one grouped card, so you can follow every thread without losing the main one.
- 18 built-in roles: implementer, tester, code-reviewer, security-reviewer, debugger, refactorer, docs-writer, researcher, planner and more
- Define your own in .dune/agents/ or with /create-subagent
- Child sessions report back to the parent, which merges the results
Each tool has a rule. Allow runs without asking. Ask puts an approval card in the chat, and the tool waits for you. Deny blocks it. Rules can differ by mode and subagent type, so a reviewer can read but never write, and nothing runs a shell in plan mode.
- Set rules in Settings ▸ Agents ▸ Permissions
- Approval cards show exactly what will run before it runs
- Hooks in .dune/hooks.json can inspect calls before and after each tool
Type @ to attach live IDE state: the semantic codebase index, your git changes, the problems panel, symbols, files or the web. Slash commands give common tasks a head start, and you can write your own as Markdown.
- @Codebase, @Git, @Problems, @Symbols, @Web, and any file
- Slash commands: /fix, /test, /docs, /code-review, /review-security, /automate and more
- Custom commands in .dune/commands/*.md
The agent checkpoints the workspace as it edits. Restore any earlier point in the conversation and the files go back with it, including files the agent created. Review the diff in the Git panel, keep what you like, and commit.
Skills are playbooks the agent follows, written as SKILL.md files you can read and edit. Memory persists across sessions and is recalled by search. While the app is idle, Dreaming runs passes that consolidate memory, refresh the code index, review past sessions and propose new skills. Every cycle can be undone.
- Skills in .dune/skills/<name>/SKILL.md, distilled from your sessions or written by hand
- Rules in .dune/rules/ (always on, or scoped by glob) and a project constitution
- Dreaming passes: memory, code index, project refresh, reflection, session digest, skill hygiene, trajectory mining
read_file, write_file, edit_file and list_files across the workspace, local or remote.
run_shell in a real terminal, plus await_job for long-running commands.
search_code for exact matches and a semantic code index with reranking for meaning.
install_packages picks the project's package manager and installs what a task needs.
web_search through Brave, DuckDuckGo, Exa or Tavily, chosen in Settings ▸ Agents.
open_browser and browser_control to load a page, click through it and read the result.
Any stdio MCP server from .dune/mcp.json, with its tools dropped straight into the loop.
Start and list tunnels to share a local dev server while the agent works.
manage_todos keeps the plan current; ask_user stops to ask you a question.
| Path | What it does |
|---|---|
| .dune/rules/ | Always-on or glob-scoped instructions |
| .dune/constitution.md | Project-wide principles the agent follows |
| .dune/skills/<name>/SKILL.md | Playbooks the agent can load |
| .dune/commands/*.md | Custom slash commands |
| .dune/modes/*.md | Custom agent modes |
| .dune/agents/ | Named subagents |
| .dune/hooks.json | Hooks that run before and after tool calls |
| .dune/mcp.json | MCP servers (stdio) |
| .dune/plans/ | Saved plan documents |
Models come through your Hypery account. You choose separate models for chat, inline completion and Cmd+K edits in Settings ▸ Agents ▸ Models. Usage is counted in AI credits.
A warm, AI-first workbench. Free to download. Bring your extensions, your themes, and your agent.