Connect an AI agent
PullMD gives agents one tool: read this URL and return clean Markdown. There are three ways to hand it to an agent. All three call your PullMD instance, so pick the one your agent supports best.
| MCP server | Claude Code skill | Prompt | |
|---|---|---|---|
| Works with | Claude Code, Cursor, Claude Desktop, claude.ai and other MCP clients | Claude Code | Any agent that can make HTTP requests |
| Install | One command or a JSON entry | Unzip into ~/.claude/skills | Paste into the system prompt |
| How the agent calls PullMD | Native tools: read_url, get_share, list_recent | curl to the REST API | curl or its own fetch tool |
| Updates | Automatic: tools live on your instance | Download the zip again | Edit the prompt |
Start with the MCP server if your agent supports MCP. The agent sees typed tools with descriptions, so it uses them without extra instructions.
Before you start
You need a running PullMD instance that the agent can reach. For an agent on your own machine, http://localhost:3000 works. For claude.ai or a teammate's agent, the instance must be on a public HTTPS address: see Installation and Reverse proxy.
Setup commands on /help
Open /help on your instance. It shows the setup commands for all three options with your instance's address already filled in.If sign-in is enabled
With PULLMD_AUTH_MODE set to single-admin or multi-user, agents must authenticate:
- Claude Code, Cursor and scripts: create an API key at
/settingsand send it asAuthorization: Bearer pmd_…. - Claude Desktop and claude.ai: their connector settings have no header field, so they use OAuth. See Connect claude.ai and Claude Desktop.