Code

Coding agents you review, change by change.

Agents work in git worktrees of their own, in the desktop app, VS Code and the CLI. Each change comes back as a diff you keep or drop, and only what you keep is written.

The desktop workbench in Code: the project's changed files on the left, src/settlement/batch.ts open in the editor, and the agent's conversation on the right with the edit it made and the tests it ran.
How it works

Agents in parallel, people in charge.

Each agent in its own worktree.

Start several sessions on one repository. Each gets a git worktree of its own, so agents never edit the same checkout, and subagents can take worktrees too.

Sessions that wait for you, and the changes waiting for approval, are listed in one place.

The workbench home: seven agent sessions marked running or waiting, and four approvals waiting for review.

Review per change.

Before a file is written, the agent shows each change as a diff. Keep or drop each one; only kept changes are written. A rule can require approval for every edit.

In VS Code the same review opens side by side in the editor.

The Mindcode panel in VS Code asking for approval of three changes to src/settlement/batch.ts, each shown as a diff with a Kept mark, and buttons to reject or apply all three.

Tab completion and inline edit.

In VS Code, Mindcode suggests the next lines as you type and rewrites a selection when you ask. Both use the fast model role, fall back to main when there is no fast model, and honour the same ignore files and permission rules as the agent.

"mindaeon.completions.enabled": true,
"mindaeon.completions.languages": ["*"]

Nothing leaves without a check.

Before a tool call sends data off your machine, its arguments are checked for credentials, private keys and personal data such as IBANs, card numbers and Turkish identity numbers. Secrets are refused; personal data asks you first.

Tool descriptions from MCP servers are pinned. A server that changes a tool later has to be approved again.

"toolFirewall": {
  "secrets": "deny",
  "personalData": "ask"
}
In the runtime

Everything a coding agent needs.

Plan, then change

Ask for a plan first and approve it before the agent edits anything.

Subagents

A session can hand work to subagents, which run under the same permissions and controls as the session.

Hooks

Run your own commands before and after tool calls, to block, check or record what an agent does.

Skills and plugins

Package instructions and tools once, and share them through your organization's library.

Commits observed, not rewritten

Commits an agent makes are recorded as the agent's. No git hooks are installed and no history is rewritten.

Any model, by role

Settings name roles such as main and fast, not models. Use Anthropic's models or an endpoint you serve yourself.

Try it on your own repository.

The desktop app, the VS Code extension and the CLI run on their own with your model key.