Purpose
Kilo-Kit is an agentic MCP runtime and curated skills catalog for AI coding assistants, with the core goal of making coding agents safer, better grounded, and more verifiable when executing tasks. It plays the role of a high-level cognitive brain (Cortex), while the host client is responsible for concrete I/O execution (Limbs), forming a division of boundaries between "cognitive reasoning" and "surgical I/O execution."
Main Capabilities
- C4 cognitive lifecycle and low-confidence escalation: All agent tasks flow through a deterministic 5-Gate state machine, and unauthorized file modifications before Gate 3 approval are blocked at the server level.
- Kilo-Sentinel supervisor and circuit breaker: Real-time middleware enforces a preflight grounding lock (editing unread files is prohibited), loop tripwires (detecting identical calls and edit churn), and records trajectories via SQLite (katl_trajectories).
- Triangular cognitive synthesis and low-confidence escalation: Combines internal SQLite memory recall, GitHub 10k+ stars patterns, and ToT DAG benchmarking (kilo_triangulate_research), automatically delegating to subagents when confidence is below 0.70.
- Fuzzy skill and alias resolver: Supports alias-based on-demand skill loading, such as brainstorming, diagnose, playwright, clean-code, tdd, grounded-research-benchmark.
- 4D quality assurance and Playwright E2E gate: Before marking a task complete, verifies Given-When-Then acceptance criteria, clean code interfaces, UI/UX aesthetics, and automated Playwright browser/DOM verification.
- Protocol-level hard gates: Enforces safety rules through JSON-RPC interceptor middleware, avoiding the prompt drift problem of traditional prompt rules (such as .cursorrules, CLAUDE.md).
- 24 all-in-one MCP tool suite: Provides a complete, self-contained execution and cognitive runtime.
- 180 curated skills catalog: Organized into 6 functional modules, supports fuzzy alias resolution, for example calling kilo_get_skill("brainstorming") automatically loads productivity/brainstorming/SKILL.md.
Usage
1. Quick install (recommended): Install globally and automatically configure all detected AI clients (Cursor, Claude, Windsurf, Antigravity, Gemini), then verify system health.
2. Zero-install NPX (direct IDE integration): No global installation required; add Kilo-Kit directly to the client's MCP configuration, supporting Cursor (.cursor/mcp.json), Claude Desktop (claude_desktop_config.json), Windsurf, and Antigravity / Gemini CLI.
3. Team repository rollout: Bootstrap the Kilo-Kit C4 cognitive protocol into a project repository (CLAUDE.md, AGENTS.md, GEMINI.md), or use the global git alias git kilo-init.
Suitable Scenarios
Suitable for developers using MCP-capable clients such as Cursor, Claude, Windsurf, Antigravity, and Gemini, when they need AI coding agents to follow stricter diagnosis, planning, and verification processes, or when they want teams to unify the coding agent cognitive protocol.
Limitations
The materials do not provide specific installation commands, configuration examples, tool name details, benchmark data, performance metrics, or pricing information; these contents are not yet provided. The project is distributed under the Apache 2.0 license.
