v1.21.0May 25, 2026
AI Agent Orchestrator
AI Agent Orchestrator. Also in this release: Mission Control, MCP Tool Activity Badges, Smarter Session Resume and MCP Settings Redesign.
Download v1.21.0
The download page always has the newest build; installers for this version are on GitHub.
What changed
New features
17- Mention any AI agent and your current agent will delegate the task — type @Codex write jokes in a Claude terminal and Claude calls Codex headlessly, captures the output, and reasons over it
- Works with every coding CLI — Claude, Codex, Gemini, Amp, OpenCode, Qwen, and Aider can all be called or do the calling
- Multi-agent workflows in natural language — ask "have Codex write jokes and score them until 8/10" and the orchestrating agent handles the loop automatically
- Run agents in parallel — mention multiple agents and the AI decides whether to run them simultaneously or sequentially based on your prompt
- New MCP tools available to every AI terminal — run_agent delegates a prompt to another agent, list_agents shows which agents are installed on your system
- Manage from Settings → MCP — toggle the Orchestrator feature on/off, see both tools with their descriptions and status
- Press the grid icon in the toolbar to open Mission Control — a full-screen overlay showing every project as a card
- Each card shows what's running — terminal count by agent type (3 Claude, 1 Codex), browser tabs, HTTP requests, and database connections at a glance
- Custom commands are parsed and badged — if a terminal runs claude --model opus, the card shows a colored Claude badge with the flags
- Active projects float to the top, inactive projects are grouped below with a subtle divider
- Keyboard navigation — arrow keys to move between cards, Enter to switch, Escape to close
- Spring-animated transitions — cards shrink in from full size on open, the selected card zooms back out on close
- Live badges appear in the terminal area when an AI agent calls an MCP tool — database queries, HTTP requests, or agent orchestration calls
- Animated status indicators — a pulsing dot while running, a checkmark or error icon when done, with elapsed time counter
- Click any badge to open a detail panel showing the full result — query output as a formatted table, HTTP response with status/headers/body, or the delegated agent's response
- Color-coded by category — green for database, blue for HTTP, amber for orchestrator
- Auto-dismiss or clear all — badges stay visible until dismissed so you can review results at your pace
Improvements
8- Session names appear in the Resume panel — Claude sessions show the AI-generated title, Codex sessions show the thread name, OpenCode sessions show the session title
- Model info displayed — see which model (Opus, Sonnet, GPT-4o) was used in each session
- Git branch shown — know which branch a session was on before resuming
- Card-style layout — sessions now show as cards with a colored accent bar matching the agent, replacing the flat row layout
- Search by session name — find sessions by their AI-generated title, not just the first prompt
- Tools tab shows every registered MCP tool grouped by feature — see tool names, descriptions, and whether they're read-only or mutating
- Feature toggle immediately changes tool visibility — no restart required
- Bridge health card shows enabled tool count and registry version
Bug fixes
3- Fixed macOS build failing during notarization step
- Fixed MCP server failing to start when the bridge wasn't fully initialized
- Fixed Agent Input overlay not recognizing built-in AI agents for chip display — Claude Code, Codex, Gemini, and all others now show as chips when mentioned