Kiro introduces “powers” to load developer expertise on demand
AI agents often stumble between two problems: missing framework context that forces repeated trial-and-error, or drowning in an overloaded context window when many tools are loaded at once. Kiro’s answer is a packaging and activation layer called powers that bundles MCP tool definitions, steering guidance, and activation rules into self-contained units that load only when relevant.
The friction powers aims to solve
Current agent toolchains fragment knowledge and tooling across MCP servers, Skills, dynamic-loading systems, and per-client rules. That fragmentation forces repeated configuration when switching tools or environments. Kiro powers are presented as a simpler developer experience: install a power once and the agent gains the bundled tooling and workflow guidance for that domain.
What a power contains
- POWER.md: a steering entry that defines activation keywords, onboarding steps, and workflow-specific steering files.
- MCP server configuration: the actual tool endpoints and connection details.
- Hooks and steering assets: optional runtime hooks, slash commands, or workflow maps loaded on demand.
Key technical benefits
- Dynamic MCP tool loading: powers activate based on keywords (for example, “database” or “checkout”), which keeps baseline context usage near zero until a specific capability is needed — addressing token bloat when many MCP servers are configured.
- One-click install and discovery: powers can be browsed and installed from the Kiro IDE or kiro.dev, with prompts for API keys or env vars on first use instead of manual JSON configuration.
- Ecosystem and sharing: partner powers (Stripe, Supabase, Figma, Netlify, Neon, Postman, and others) and community-built or private powers can be imported from GitHub or local repos.
- Cross-compatibility (coming soon): a plan to make powers work across tools that implement the Model Context Protocol so a single POWER.md could serve multiple agent platforms.
A practical angle
Powers aim to keep agent context focused by loading workflow-specific steering documents only when those workflows are active (for example, RLS policies vs. Edge Functions for Supabase). The model emphasizes continual expansion of agent expertise as new powers are installed, rather than preloading every capability.
Read the original Kiro post for examples and installation details: Introducing powers — Kiro