Manu's Vault

Kiro-Powers

4 min readStatus:Seedling

Kiro Powers

Kiro Powers are dynamic capability bundles that combine Model-Context-Protocol tools with domain-specific framework expertise, loaded on demand rather than at session start. They solve the context overflow problem that static MCP configurations create in AI coding agents.

The Problem They Solve

Traditional MCP setups load all connected servers upfront. Five servers can consume 50,000+ tokens — up to 40% of the context window — before a single prompt is processed. This is a direct instance of Context-Rot: context fill degrading agent quality before work begins. Perplexity CTO Denis Yarats cited this 40–50% context window consumption as a primary reason his team moved away from MCP in production (Ask 2026 conference, March 2026).

Powers address this by keeping baseline context usage near zero. Install five powers; consume nothing until the agent detects a relevant trigger in the conversation.

The Three Power Components

Every Kiro Power is packaged as three interconnected parts:

  • POWER.md — The entry-point steering file. Documents available MCP tools and defines the usage scenarios that trigger activation. The agent reads this file to decide whether to load the power.
  • MCP Server Configuration — Connection details and full tool definitions for the underlying MCP server.
  • Steering Files / Hooks — Workflow-specific guidance and automated actions, triggered by slash commands or lifecycle events within the conversation.

This structure is an application of Progressive-Disclosure-Context: only POWER.md metadata loads at startup; full tool definitions and steering only activate when relevant.

The Activation Mechanism

Powers activate via keyword detection. Mentioning “database” loads the Supabase or Neon power’s tools and steering guides. Switching topics triggers deactivation of the previous power and (optionally) activation of another. This dynamic load/unload cycle means the context window reflects only the current task’s tooling, not the entire capability stack.

The Supabase power example demonstrates workflow-aware loading: different documentation loads based on what the developer is doing — Row Level Security policies vs. Edge Functions vs. schema migration — without loading all patterns simultaneously.

Ecosystem and Cross-Compatibility

Launch partners include Datadog, Dynatrace, Figma, Neon, Netlify, Postman, Stripe, Supabase, and Strands Agent. Community-built powers exist for AWS CDK and Amazon Aurora DSQL.

Powers are currently available in Kiro IDE and are positioned for cross-compatibility with Kiro CLI, Cline, Cursor, Claude Code, and other tools. The vision is that Model-Context-Protocol provides the standard for tool communication while powers extend it with packaging, activation, and knowledge transfer standards — analogous to how Agent-Skills package domain expertise in Claude’s ecosystem.

Sources

  • Kiro (2026). “Introducing Powers.” kiro.dev. Retrieved March 2026. https://kiro.dev/blog/introducing-powers/

    • Primary source; defines powers as MCP + framework expertise bundles; 50,000-token context overflow example; three-component anatomy; keyword-based activation; launch partner ecosystem; cross-tool compatibility vision

  • Sinha, Aakash et al. (2025). “MCP Tool Descriptions Are Smelly! Towards Improving AI Agent Efficiency with Augmented MCP Tool Descriptions.” arXiv preprint, arXiv:2602.14878. https://arxiv.org/html/2602.14878

    • Academic research on MCP tool description inefficiency; establishes evidence base for the context cost problem that Powers address

  • Spring Team (2025). “Dynamic Tool Updates in Spring AI’s Model Context Protocol.” Spring.io Blog. https://spring.io/blog/2025/05/04/spring-ai-dynamic-tool-updates-with-mcp/

    • Practitioner implementation of dynamic tool loading in MCP; demonstrates use cases including feature flags, permission-based loading, and runtime plugin systems that parallel the Powers pattern

  • Microsoft Azure Dev Community (2025). “MCP vs mcp-cli: Dynamic Tool Discovery for Token-Efficient AI Agents.” Microsoft Tech Community. https://techcommunity.microsoft.com/blog/azuredevcommunityblog/mcp-vs-mcp-cli-dynamic-tool-discovery-for-token-efficient-ai-agents/4494272

    • Industry practitioner perspective on dynamic vs. static tool discovery; token efficiency framing; on-demand tool definition loading as an alternative to upfront loading

  • Lushbinary (2026). “AI Coding Agents 2026: Claude Code vs Antigravity vs Codex vs Cursor vs Kiro vs Copilot vs Windsurf — Pricing & Features Compared.” Lushbinary.com. https://lushbinary.com/blog/ai-coding-agents-comparison-cursor-windsurf-claude-copilot-kiro-2026/

    • Ecosystem comparison; positions Kiro Powers as a differentiating feature in the AI IDE landscape; context on how the broader tooling ecosystem approaches dynamic capability loading

Note

This content was drafted with assistance from AI tools for research, organization, and initial content generation. All final content has been reviewed, fact-checked, and edited by the author to ensure accuracy and alignment with the author’s intentions and perspective.

Graph View

Backlinks