A coding agent that lives where you already work
Differential-rendering TUI, multi-mode authentication (Claude Pro, ChatGPT Plus, 15+ API keys, or offline Ollama), workspace memory, and atomic code mutations.
The 6 Pillars of the Agent Harness
The operational runtime scaffolding that elevates raw intelligence into an autonomous software engineer.
Loop
The central execution runtime. An agent harness runs a continuous loop that observes the environment, plans an action, executes it, and evaluates the result — repeating until the task is done.
Terminal intelligence engineered for developer flow
Engineered for speed, privacy, and full developer agency. Explore how Cube orchestrates models, workspace context, and safe execution directly inside your shell.
Multi-Gateway Model Matrix
Connect directly to Claude Pro, ChatGPT Plus, or Grok memberships via browser PKCE OAuth 2.0 with zero token markup. Plug in any of 15+ developer API keys, or run 100% offline with local Ollama models.
- ›PKCE OAuth for Claude Pro / ChatGPT Plus / Grok
- ›15+ Developer API providers with unified fallback routing
- ›100% Air-gapped offline inference via Ollama & LM Studio
Workspace Skill & Rule Discovery
Cube crawls project directories and walks up parent folders to automatically discover AGENTS.md rules, repository guidelines, and custom skill scripts. Injected into every turn with zero prompt copy-pasting.
- ›Automatic AGENTS.md rule discovery across parent trees
- ›Dynamic skill loading with isolated execution runtimes
- ›Zero prompt maintenance across multiple monorepos
Human-in-the-Loop Safeguards
No silent overwrites or rogue actions. Before applying atomic disk modifications or running potentially destructive terminal commands, Cube prompts for explicit developer confirmation with full unified diff previews.
- ›Human confirmation required for sensitive bash commands
- ›Side-by-side colorized unified diff previews before write
- ›One-key rollback and atomic disk commit guarantees
Interactive Intent Clarification
When requirements are underspecified or architectural tradeoffs arise, Cube presents keyboard-driven multiple-choice questions right in your shell to lock down exact implementation specs.
- ›Interactive arrow-key multiple choice in terminal TUI
- ›Clarify underspecified requirements early in planning
- ›Instant architectural alignment without endless chat loops
Model Context Protocol Foundation
Built-in Model Context Protocol (MCP) client architecture. Seamlessly connect external tool servers, database inspectors, browser automation runtimes, and proprietary enterprise endpoints through open standards.
- ›Universal MCP client supporting stdio and SSE transports
- ›Connect databases, GitHub, browser automation, and APIs
- ›Community server ecosystem with zero custom adapter code
Parallel Subagent Task Delegation
Deconstruct massive refactoring projects into isolated parallel subagent workers. Subagents conduct deep codebase exploration, write tests, and apply scoped edits concurrently without blocking your main prompt loop.
- ›Spawn background worker subagents with isolated context
- ›Parallel codebase exploration, refactoring, and test writing
- ›Automatic dependency resolution and unified pull review
Multi-gateway model routing & OAuth
Multi-gateway model routing & OAuth
Multi-Gateway Model Matrix
Connect directly to Claude Pro, ChatGPT Plus, or Grok memberships via browser PKCE OAuth 2.0 with zero token markup. Plug in any of 15+ developer API keys, or run 100% offline with local Ollama models.
- ›PKCE OAuth for Claude Pro / ChatGPT Plus / Grok
- ›15+ Developer API providers with unified fallback routing
- ›100% Air-gapped offline inference via Ollama & LM Studio
Command-line precision built for flow
Keep your fingers on home row. Instant slash commands give you total control over models, memory compression, and thread branches without interrupting your work.
/model [model_name]Switch the active reasoning model instantly without restarting your agent session.
Switched active model: From : claude-3-5-sonnet-20241022 To : claude-3-7-sonnet-20250219 (Thinking enabled) Context preserved across 8 turns.
Frequently Asked Questions
Architecture, data privacy, subscriptions, and getting started.