Skip to main content
Use this page for the agent-specific layer: how an AI workflow should choose a source, what governance to apply, and when to stop. Choose an interface contains the full setup matrix for REST, WebSocket, OpenAPI, SDKs, CLI, MCP Server, Skill, and Markdown. Do not start with a host name. Start with the action: cite docs, generate a route, run a shell command, call an MCP tool, use a reusable Skill, edit application code, or stream data. Host and harness labels only matter after the workflow selects a surface the environment can actually use.

​
Source Hierarchy

​
Choosing A Surface

1

Name the host

Identify whether the user is in a coding-agent host, OpenClaw, a web LLM, CI, or application code.
2

Name the execution ability

Decide whether the environment can call tools, run shell commands, edit code, install a skill, read Markdown, or only answer from context.
3

Pick the narrowest surface

Use Markdown and OpenAPI for route decisions, CLI for terminal calls, MCP for tool-calling hosts, Skill for reusable procedures, and SDKs for application code.
4

Add the data gate

Any generated backtest, alert, export, dashboard, or model job should include data-quality checks.

​
Agent Safety Rule

Do not call MCP, CLI, SDKs, OpenAPI, Markdown, llms.txt, or Skills “agents.” They are surfaces. Do not call OpenClaw a Skill or MCP server. It is a harness that can use those surfaces when configured. This distinction prevents setup docs from giving a user the wrong install path.

​
Stop And Ask When

Ask whether the user means Hyperliquid core, Hyperliquid Spot, HIP-3, HIP-4, or Lighter before generating a route.
Do not give MCP setup to a host that cannot call MCP tools, or Skill setup to a host that cannot load SKILL.md workflows.
Treat the request as outside the current public API contract until a public route exists.
Add data-quality status, coverage, latency, or incidents before the agent writes a backtest, alert, export, dashboard, or model job.

​
Next Step

Use Coding agents for host-level behavior, MCP server for MCP setup, Skill for SKILL.md workflows, and AI clients for prompt patterns.
Last modified on August 13, 2026