Aider gets a cockpit: worktrees, approval gates, and message surgery

What it does AiderDesk is a TypeScript/Electron desktop app built on top of the Aider CLI — what every GUI aspires to be when it grows up: an “orchestration layer.” It organizes work into Projects and Tasks, and runs each task in its own Git worktree so the AI can refactor freely without ever touching your active branch. Every proposed change lands in a diff viewer (side-by-side or unified) before it hits disk, and shell commands or file operations can be gated behind explicit human approval. It talks to 25+ model providers and can delegate multi-step work to subagents you define via Agent Profiles.

The interesting bit The control extends into the conversation itself: you can delete individual messages from chat history to scrub a bad context loop, or fork a task to explore an alternative implementation without losing the original. It also speaks MCP in both directions — consuming external MCP servers for things like Jira or internal wikis, and exposing itself as an MCP server so Claude Desktop or Cursor can drive it.

Key highlights

• Per-task Git worktrees with a built-in merge workflow back into your main flow
• Approval gates for destructive actions; line-by-line diff review before anything is applied
• Message-level history editing, plus task duplication and forking
• Context engine on LanceDB vector embeddings with pinnable docs and code symbols; Skills load on demand to keep token usage lean
• Extension system with 30+ lifecycle hooks, custom TypeScript tools exposed to the AI, and injectable React components — plus a REST API and IDE connectors for IntelliJ and VS Code

Caveats

• It’s an Electron app layered over a CLI — if you live in the terminal, this is a GUI where you already had one, albeit a far richer one.
• The README doesn’t explain how the LanceDB memory engine divides duties with Aider’s own repo-map; the context pipeline is underspecified here.
• The capability list is long, and a few items (voice control, “top-tier voice models”) get a single sentence and no detail.

Verdict Worth a look if you like Aider’s output but want review gates, parallel isolated tasks, and a dashboard for token spend. If raw Aider in a terminal already fits your habits, this adds a lot of app to a workflow that was doing fine.