You rarely need to rebuild your agent to give it a new ability. IKKF is a skill router that lets your running agents discover and load new tools on-demand, based on the conversation context.
Bundling dozens of rarely-used tools into every agent deployment creates bloat and slows iteration. IKKF was built to separate skill storage from your core agent logic. Your agent can request the exact code it needs, only when a user's request matches a skill's description.
This is a fork-first, open-source project. You deploy and control your own instance.
- Fork this repository to your GitHub account.
- Deploy your fork to Cloudflare Workers:
npm install npx wrangler deploy
- Add Skills: Drop
.jsor.tsfiles into the/skillsfolder in your fork. Each file must export a default function. Push to deploy.
Your agents can now call your IKKF endpoint to discover and load these skills dynamically. Use npm run dev for local testing.
- On-Demand Loading: Skills are fetched by your agent only when their description matches the user's intent.
- Edge Runtime: Deploys globally on Cloudflare Workers. Core skill routing logic typically completes in under 100ms.
- Zero Dependencies: The runtime has no npm dependencies, reducing supply-chain risk.
- Skill Isolation: Skills run in separate, isolated environments. A broken skill does not crash the router.
- Live Updates: Add, update, or remove skills by committing to your
/skillsfolder. No agent redeploys needed. - Centralized Skills: A single IKKF fleet can serve multiple, different agents.
- Your agent receives a user message (e.g., "Check the rain tomorrow").
- The agent calls your IKKF endpoint with this message.
- IKKF evaluates all skills in its
/skillsfolder, matching the message against each skill's description. - IKKF returns the code for the highest-matching skill(s) to your agent.
- Your agent executes the returned code.
This is not a proxy or middleware. IKKF only returns the skill's source code to your agent, which then runs it.
Skill files are currently limited to 1KB in size. This enforces focused, single-purpose skills but means you cannot bundle large libraries or complex logic directly into a skill file.
You can inspect a live instance: https://the-fleet.casey-digennaro.workers.dev
MIT