Reference
Skill files
Two markdown files that teach any AI agent to operate AffiliateOS — no MCP client required. Paste a URL into your agent's context and it knows the tools, the loop, and the failure modes.
The files
| File | What it teaches |
|---|---|
| /skill/SKILL.md | The full operating manual: connecting, authentication, all 16 MCP tools with inputs and outputs, the discover → mint → measure loop, REST fallback endpoints, error conventions, and network quirks. |
| /skill/HEARTBEAT.md | A scheduled monitoring routine: poll check_events with a high-water mark, triage commissions and reversals, catch broken network connections, and report deltas to the human — cheap enough to run hourly. |
Using them
With Claude Code, drop them into a project as a skill or just reference the URL:
mkdir -p .claude/skills/affiliateos
curl -o .claude/skills/affiliateos/SKILL.md https://affiliateos.dev/skill/SKILL.md
curl -o .claude/skills/affiliateos/HEARTBEAT.md https://affiliateos.dev/skill/HEARTBEAT.mdAny other agent framework works the same way — the files are plain markdown with no front-matter requirements. An agent holding SKILL.md can drive the REST API directly with an API key even if it cannot speak MCP.
Skill files vs the MCP server
They complement each other. The MCP server gives an agent live, schema-validated tools; the skill files give it the judgment layer — which tool to reach for, what the commission lifecycle means, when a reversal is noise versus a pattern. Connect MCP for the hands, load the skill file for the head.
curl always gets documentation that matches the deployed tool surface.