Pump AI Agents

AI Agent Integration — Agent Scaffolding & Discovery

Configure AI agents (GitHub Copilot and Gemini) to work effectively with the Pump SDK through instruction files, skill registries, and MCP server integration.

Agent Instruction Files

.well-known Discovery Files

LLM Context Documents

Skills Registry Schema

{
    "schema_version": "v1",
    "name": "pump-fun-sdk",
    "skills": [
        {
            "id": "pump-sdk-core",
            "name": "Pump SDK Core",
            "description": "...",
            "path": ".github/skills/pump-sdk-core.skill.md",
            "tags": ["typescript", "solana", "sdk"]
        }
    ]
}

Terminal Management (Agent Rules)

• Always use background terminals (isBackground: true) for every command
• Always kill the terminal after completion — never leave terminals open
• Do not reuse foreground shell sessions
• If a terminal appears unresponsive, kill it and create a new one

Security Invariants for Agents

• ONLY official Solana Labs crypto: solana-sdk, @solana/web3.js, solana-keygen
• Zeroize all key material after use
• File permissions 0600 for keypairs
• No network calls for key generation
• All amounts use BN (bn.js) — never JavaScript number for financial math
• createInstruction (v1) is deprecated — use createV2Instruction

MCP Server Integration

{
    "mcpServers": {
        "solana-wallet": {
            "command": "npx",
            "args": ["-y", "@pump-fun/mcp-server"],
            "transportType": "stdio"
        }
    }
}

Patterns to Follow

• Keep agent instruction files concise and actionable
• Point agents to specific skill files for domain knowledge
• Update skills.json when adding or modifying skill files
• Maintain consistency between AGENTS.md, COPILOT.md, and GEMINI.md

Common Pitfalls

• Agent instruction files that are too long get truncated or ignored
• Missing skills from skills.json means agents won't discover them
• Terminal management rules are critical in Codespaces — stale terminals block operations
• MCP server must be registered in the agent's tool configuration