Ask Claude about pump.fun launches: MCP setup for Claude Desktop and Claude Code

To ask Claude about pump.fun launches, add the Anaxer MCP server to Claude Desktop or Claude Code - one config block running npx -y @anaxer/mcp with your API key. Claude then calls typed tools for launches, graduations, trades, prices, and launchpad stats, and answers with analysis of real on-chain events.

What can Claude do with pump.fun data?

Language models are good at exactly the part of pump.fun research that is tedious for humans: reading dozens of structured events, applying consistent criteria, and explaining what stands out. Once Claude can call real data tools, questions like these become one-message answers:

• "What launched in the last hour, and which launches have a website and Twitter?"
• "Which tokens graduated today, and how fast did each fill its bonding curve?"
• "Is launch volume today unusual compared to the past week?"

Without a data connection Claude will invent plausible-looking mints and prices. The MCP server closes that gap - and because every tool is read-only, connecting it risks nothing.

How do I set up Claude Desktop?

Edit claude_desktop_config.json (Claude → Settings → Developer → Edit Config) and add the server. You need Node.js 20 or newer installed, and an API key from the console - free to create, no card.

{
  "mcpServers": {
    "anaxer": {
      "command": "npx",
      "args": ["-y", "@anaxer/mcp"],
      "env": {
        "ANAXER_API_KEY": "YOUR_KEY"
      }
    }
  }
}

Restart Claude Desktop. The tools indicator in the chat input now lists the nine anaxer tools - the full reference is in the MCP server docs.

How do I set up Claude Code?

Claude Code uses the same mcpServers shape in its MCP config, so the block above works unchanged. The difference is what you can do with it: in Claude Code, the same session that queries live data can also write and run code against it, and project instructions can carry your screening criteria between sessions.

What analysis prompts work best?

Start with single-tool questions, then move to the comparisons where Claude earns its keep.

Launch screening - built on list_creations:

"List the last 30 pump.fun launches. Group them into: has socials, Mayhem mode, and bare. Flag anything where the same creator appears twice."

Graduation analysis - built on list_graduations:

"Show me today's graduations. For each, report time-to-graduate and liquidity moved. Which graduated unusually fast, and is fast graduation correlated with more liquidity in this sample?"

Graduation events carry timeToGraduateSeconds and liquiditySol, so Claude computes these comparisons from real fields rather than estimates. If graduations are new to you, what a graduation is and how to detect one covers the mechanics.

Market pulse - built on get_launchpad_stats:

"Get launchpad stats for the past 24 hours and the 24 hours before that. Is creation volume accelerating? Summarize in three sentences."

Live sample - built on tail_stream:

"Tail the creations stream for 20 seconds and tell me what launched while we were talking."

Each tail is a bounded collect - up to 100 events or 30 seconds, whichever comes first - so Claude gets a live sample without holding a connection open.

Troubleshooting: why can't Claude see the tools?

• Tools missing after editing the config. Claude Desktop only reads claude_desktop_config.json on startup - restart the app fully. Validate the JSON: a trailing comma silently breaks the whole mcpServers block.
• Server fails to launch. Check node --version is 20+, and that ANAXER_API_KEY is set in the env block - the server exits with a stderr message if the key is missing.
• unauthorized errors. On REST tools, fix the key and ask again. On tail_stream, a bad key kills the shared WebSocket - fix the key and restart Claude so the server relaunches.
• subscription_limit on tail_stream. Each open tail holds one plan subscription; the free plan allows one concurrent stream. Let the running tail finish before starting another.

Where does this fit in a bigger system?

Claude-plus-MCP is the research and judgment layer: screening launches, comparing graduation cohorts, and pressure-testing strategy ideas against live data. It is deliberately not the execution layer - the tools cannot sign or trade - and it is not the always-on layer either, since analysis happens when you ask. The Solana trading agent guide shows how all three layers fit together, and when a strategy needs a persistent feed, @anaxer/sdk runs the always-on half with the same event shapes Claude has been reasoning over. If your editor is Cursor rather than Claude, the Cursor walkthrough mirrors this setup.