AI Agent Integration — Agent Scaffolding & Discovery
Configure AI agents (GitHub Copilot, Claude, Gemini) to work effectively with the Pump SDK through instruction files, skill registries, and MCP server integration.
Agent Instruction Files
| File |
Agent |
Purpose |
AGENTS.md |
Universal |
Project overview, architecture, security rules |
CLAUDE.md |
Claude Code |
Key patterns, critical rules |
COPILOT.md |
GitHub Copilot |
Instructions pointer |
GEMINI.md |
Google Gemini |
Instructions pointer |
.github/copilot-instructions.md |
Copilot Chat |
SDK patterns, security rules |
.well-known Discovery Files
| File |
Purpose |
.well-known/ai-plugin.json |
AI plugin manifest |
.well-known/agent.json |
Agent capabilities and config |
.well-known/skills.json |
Skills registry (15+ skills) |
.well-known/security.txt |
Security contact information |
LLM Context Documents
| File |
Size |
Purpose |
llms.txt |
Short |
Quick reference for LLMs |
llms-full.txt |
Long |
Comprehensive LLM context |
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"],
"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, CLAUDE.md, and COPILOT.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
1---2name: pump-ai-agents3description: AI agent integration layer for the Pump SDK — agent instruction files, .well-known discovery, LLM context documents, 15+ skill files, MCP server prompts, and terminal management rules for GitHub Copilot, Claude, and Gemini.4---56# AI Agent Integration — Agent Scaffolding & Discovery78Configure AI agents (GitHub Copilot, Claude, Gemini) to work effectively with the Pump SDK through instruction files, skill registries, and MCP server integration.910## Agent Instruction Files1112| File | Agent | Purpose |13|------|-------|---------|14| `AGENTS.md` | Universal | Project overview, architecture, security rules |15| `CLAUDE.md` | Claude Code | Key patterns, critical rules |16| `COPILOT.md` | GitHub Copilot | Instructions pointer |17| `GEMINI.md` | Google Gemini | Instructions pointer |18| `.github/copilot-instructions.md` | Copilot Chat | SDK patterns, security rules |1920## .well-known Discovery Files2122| File | Purpose |23|------|---------|24| `.well-known/ai-plugin.json` | AI plugin manifest |25| `.well-known/agent.json` | Agent capabilities and config |26| `.well-known/skills.json` | Skills registry (15+ skills) |27| `.well-known/security.txt` | Security contact information |2829## LLM Context Documents3031| File | Size | Purpose |32|------|------|---------|33| `llms.txt` | Short | Quick reference for LLMs |34| `llms-full.txt` | Long | Comprehensive LLM context |3536## Skills Registry Schema3738```json39{40 "schema_version": "v1",41 "name": "pump-fun-sdk",42 "skills": [43 {44 "id": "pump-sdk-core",45 "name": "Pump SDK Core",46 "description": "...",47 "path": ".github/skills/pump-sdk-core.skill.md",48 "tags": ["typescript", "solana", "sdk"]49 }50 ]51}52```5354## Terminal Management (Agent Rules)5556- **Always use background terminals** (`isBackground: true`) for every command57- **Always kill the terminal** after completion — never leave terminals open58- Do not reuse foreground shell sessions59- If a terminal appears unresponsive, kill it and create a new one6061## Security Invariants for Agents6263- ONLY official Solana Labs crypto: `solana-sdk`, `@solana/web3.js`, `solana-keygen`64- Zeroize all key material after use65- File permissions `0600` for keypairs66- No network calls for key generation67- All amounts use `BN` (bn.js) — never JavaScript `number` for financial math68- `createInstruction` (v1) is deprecated — use `createV2Instruction`6970## MCP Server Integration7172```json73{74 "mcpServers": {75 "solana-wallet": {76 "command": "npx",77 "args": ["-y", "pump-fun-mcp"],78 "transportType": "stdio"79 }80 }81}82```8384## Patterns to Follow8586- Keep agent instruction files concise and actionable87- Point agents to specific skill files for domain knowledge88- Update `skills.json` when adding or modifying skill files89- Maintain consistency between `AGENTS.md`, `CLAUDE.md`, and `COPILOT.md`9091## Common Pitfalls9293- Agent instruction files that are too long get truncated or ignored94- Missing skills from `skills.json` means agents won't discover them95- Terminal management rules are critical in Codespaces — stale terminals block operations96- MCP server must be registered in the agent's tool configuration97
nirholas