Features

Tool Calling

The agent reads files, writes code, executes shell commands, runs code in 6 languages, searches your codebase with ripgrep, manages git, generates images and video, searches the web, and spawns up to 6 sub-agents (4 profiles — code analysis, web research, reasoning coder, verifier) for parallel research.

Permission System

Three-tier safety model. Whitelisted tools (read, search, generate) execute instantly. Graylisted tools (write, execute, delete) show an approval modal. Blacklisted operations (rm -rf /, /dev access, fork bombs) are hard-blocked — no override possible.

Memory & Knowledge

Persistent memory saves facts across sessions (global + per-session scopes). A SQLite-backed knowledge store indexes conversations by category and project. The agent builds context over time — it remembers what you've worked on.

Profiles & Skills

Create agent profiles with custom system prompts, model selection, tool presets, voice configuration, daily budgets, and linked document collections. Switch between a coding assistant, writing partner, and research agent in one click.

Mission Control

Visual multi-agent orchestration. Describe a goal, the AI architect plans a task DAG, you review and edit in a canvas editor, then execute with up to 8 parallel workers. Each mission gets its own git branch with auto-checkpoints.

• Architect mode or direct CSV/JSON import
• Per-task model assignment (xAI, Anthropic, OpenAI, Google, DeepSeek)
• Drag-and-drop visual DAG with connection ports
• Real-time cost tracking and pause/cancel

Workspace

Split-screen dual-agent mode. The planning agent (left) has read-only tools and designs the approach. The execution agent (right) has full access and implements the plan. A shared task board coordinates between them.

• Planning agent: restricted tools, approach design
• Execution agent: full tool access, implementation
• Context handoff with Cmd+Shift+Right
• Adjustable split ratio

Research Pipeline

Multi-agent research system with a 4-phase pipeline. The AI plans a checklist, agents search with Brave, parallel workers synthesize findings, then a final agent assembles the document. Export to Markdown with metadata.

• Plan → Search → Synthesize → Assemble
• 1–10 parallel synthesis workers
• Real-time progress with activity timeline
• Multi-provider support (Anthropic + xAI)

Worktree Arena

Parallel agent races in isolated git worktrees. Pit multiple models against each other on the same task, compare outputs side by side, run automated verification, and merge the winning approach.

• Race mode: same prompt, different models
• Isolation mode: different prompts per worktree
• AI review scoring and diff comparison
• Accept winner and merge with optional squash