Architecture

Camel-Kit is built on a 4-layer architecture designed for AI agent composability, cross-agent portability, and token efficiency. In the agent-run integration workflow, “the prompt is the product”: generated instructions own the integration behavior. The CLI still supplies compiled initialization, configuration, helper commands, and graph infrastructure; Ship is the deliberate controller-owned workflow exception described below.

Four Layers

Graph CLI has 15 subcommands:

• Analysis: stats, find, neighbors, path, subgraph
• Camel-specific: route-flow, impact, route-topology, dead-code
• Context: project-norms, project-context, route-context, migration-context
• Output: generate, visualize

Learn more about graph intelligence →

The Prompt Is the Product

Camel-Kit’s agent-run integration workflow embodies a key principle: the prompt is the product. Its behavior is primarily delivered through:

• Markdown guides that instruct AI agents how to generate code
• MCP tool definitions for real-time verification
• Graph parsers for code analysis (optional)

Compiled CLI infrastructure handles initialization, configuration, doctor, doc, nextId, and graph commands; it supports the prompt-owned workflow rather than replacing it.

One deliberate exception: the Ship workflow. /camel-ship is a short delegate to the registered camel-kit ship command, and the workflow controller — stages, run state, oversight, evidence, and guarded publication — is compiled code in the Camel-Kit CLI, not a prompt. Every other routed command remains prompt-owned.

This means:

• Cross-target portability — Skills are generated for the supported AI targets
• Easy customization — Edit Markdown files to change behavior
• Version-configurable — Distribution pins and live catalog queries control versions; publishing new bundled defaults requires a new build and project re-initialization
• Transparent — Users can read the exact instructions agents follow

Progressive Disclosure

Skills use progressive disclosure to minimize token usage:

1. Metadata (always loaded) — Skill name, description, and trigger patterns
2. SKILL.md (on trigger) — Main skill logic, loaded only when invoked
3. Guides (as needed) — Shared utilities, loaded only when referenced

Example flow:

User: "Create a Camel project for order processing"
  → Agent loads /camel-start
  → /camel-start routes new work to /camel-brainstorm
  → /camel-brainstorm loads only its required interview guides
  → Agent produces a design specification

This progressive loading keeps context usage minimal while maintaining full catalog coverage.

Context Efficiency

By combining progressive disclosure and MCP on-demand queries, Camel-Kit avoids loading full component catalogs into the agent’s context. This enables:

• Faster agent responses — less context to process per turn
• Support for smaller models — fits within constrained context windows
• Cost reduction — fewer input tokens per request
• Full coverage — every component is still verifiable via MCP

Next Steps

• Skills System — Skill tiers, progressive loading, and agent generators
• MCP Integration — Camel, Knowledge, and Citrus verification
• Forage Catalog — Configuration-driven infrastructure beans
• Graph Intelligence — Property graph analysis with 9 parsers
• Environment-in-the-Loop — How the execution environment drives code refinement