Skip to content

kcp-triage

LLM driven triage system for analyzing web pages and generating agentic clients based on KCP structure

Field Value
GitHub https://github.com/Cantara/kcp-triage
Language TypeScript
Stars 1
Last updated 2026-04-23

README

kcp-triage

Automatic agentic web service discovery — builds KCP knowledge about web services so LLM agents can interact with them.

This is an LLM-driven project. The recommended way to work with this codebase is to point your LLM agent (Claude Code, Cursor, etc.) at the repo and let it drive. The CLAUDE.md, skills, and docs are designed as agent context — not just human documentation. If you're reading this as a human, the quickest path is to plug in your agent and say "triage example.com".

Two-layer architecture

Builder (src/)

The CLI pipeline that crawls, classifies, audits, and reports on websites. It's the factory — you run it once per target site and it produces structured knowledge. The builder delegates LLM work across model tiers (Opus for orchestration, Sonnet for analysis, Haiku for grunt work) to minimize cost.

Sites (sites/<domain>/)

Each triaged site gets its own agent-ready workbench: a CLAUDE.md, skills, API inventory, unknown-endpoints list, and a KCP knowledge.yaml manifest. These aren't just JSON blobs — they're project scaffolds that an LLM agent can load and immediately use to interact with the target service.

Builder (src/)                        Sites (sites/<domain>/)
┌──────────────────────┐              ┌──────────────────────┐
│ crawl → classify →   │    output    │ CLAUDE.md            │
│ audit → synthesize → │ ──────────► │ skills/              │
│ generate → manifest  │              │ apis/, unknowns/     │
└──────────────────────┘              │ knowledge.yaml       │
   run once per site                  └──────────────────────┘
                                         load into your agent

Quick start

bun install
export ANTHROPIC_API_KEY=sk-ant-...

# Triage a site end-to-end
just triage https://example.com

# Or step by step
just init https://example.com
just scan example.com
just report example.com

Development

This project uses Spec-Driven Development. New features go through: Issue → Branch → Spec → Schema → Implement → Skill → Test → PR.

just check       # TypeScript check
just test        # Run tests
just fmt         # Biome format
just lint        # Biome lint
just sites       # List all triaged sites

See CLAUDE.md for full conventions, skills/ for agent-consumable workflow guides, and just --list for all available recipes.