GMGN Skills, CLI and MCP (2026): The Official AI Agent Toolkit

gmgn skills is the official open-source toolkit from GMGN for building AI trading agents. Published at github.com/GMGNAI/gmgn-skills under the MIT license, it wraps the GMGN Agent API into 12 discrete skills that AI coding tools such as Claude Code, Cursor, Cline, Codex and OpenCode can call directly. Install it with npx skills add GMGNAI/gmgn-skills, or install the standalone command-line tool with npm install -g gmgn-cli.

If you have searched for "gmgn mcp", "gmgn cli", "gmgn github" or "gmgn vibe coding kit", this is the thing all of those point at. It is the reference implementation of the API, the fastest way to get an agent reading GMGN data, and, with an API key, a way to place real trades from a chat prompt. That last part is where the caution comes in.

This guide covers what the package contains, how to install and configure it in each IDE, how the MCP-style architecture works, what natural-language tasks it handles well, and how to run it without handing a language model unsupervised access to your money. It is an independent guide; GMGN AI App Guide is not affiliated with GMGN. The underlying API is described in the GMGN API overview.

What is gmgn-skills?

GMGN's web terminal, the GMGN AI app, is built for humans clicking fast. gmgn-skills is built for agents. Instead of a monolithic SDK, it follows the "skills" convention that AI coding tools adopted in 2025–2026: each capability is a self-contained folder with a description the agent reads, plus the code that executes it. The agent picks the skill that matches your request, fills in the parameters from your prompt, and runs it.

Key facts from the repository as of September 2026:

The star count is modest by crypto standards, which is itself useful context: this is a young, actively changing package, not a mature SDK. Read the README on every upgrade.

GMGN on GitHub: official versus everything else

Searching "gmgn github" or "gmgn bot github" returns dozens of repositories. Only the GMGNAI organization is official. Everything else falls into three buckets: community wrappers around the old scraped endpoints (mostly broken), unrelated projects that happen to use the name, and outright malware disguised as "gmgn bot" or "gmgn sniper" downloads that ask for your private key or seed phrase. GMGN never asks for either, and its hosted wallets do not even allow key export. If a repo asks for a seed phrase, close the tab. More on that in the phishing guide.

The 12 GMGN skills

The package groups its skills roughly along the same lines as the API: read data, analyze, act.

The exact naming of the twelfth entry varies with releases; the README is authoritative. What matters is the split: nine or ten read-only skills, and two that move money (orders and multi-wallet trading). Your safety configuration should treat those two differently from the rest.

How to install gmgn-skills and gmgn-cli

There are two install paths. They can coexist.

Option A: add the skills to an AI IDE

npx skills add GMGNAI/gmgn-skills

This uses the skills installer to copy the skill definitions into the location your agent tool reads. After installation, the agent can discover the skills by name and description.

Option B: install the standalone GMGN CLI

npm install -g gmgn-cli

This gives you a gmgn command you can run from a terminal or from any agent that can execute shell commands. It is the better choice for scripts, cron jobs and CI, and for agents that do not support the skills format natively.

Both paths require Node.js and, for anything beyond the shared demo key, a GMGN API key. Follow how to get a GMGN API key to generate the Ed25519 pair and bind it at gmgn.ai/ai, then supply the key and private key path through the configuration method the README describes (typically environment variables or a config file). Never paste the private key into a chat prompt; the agent's transcript is not a safe place for it.

Setting up gmgn skills in Claude Code, Cursor, Cline and Codex

Each tool loads skills slightly differently. The pattern is the same: install, configure credentials outside the prompt, restart the agent, and confirm it lists the GMGN skills.

1. Claude Code. Run the npx skills add command from your project directory. Claude Code reads skill folders from its skills locations and surfaces them by name. Ask it "list the GMGN skills you have" to confirm. Put GMGN_API_KEY and the key file path in your shell environment or a .env the agent is told to read, and add the key file to .gitignore.
2. Cursor. Install the same way, then confirm the skills appear in the agent's tool list. Cursor's agent mode can execute the skill scripts; enable "ask before running commands" so order placement pauses for approval.
3. Cline. Cline runs as a VS Code extension with explicit approval for each command by default. Keep that default on for any skill in the write group.
4. Codex. Codex CLI reads project-level instructions and can run shell commands. Install gmgn-cli globally and reference it in your project's agent instructions so Codex knows the gmgn command exists and what it is for.
5. OpenCode. Same approach as Codex: global gmgn-cli install plus project instructions.

If your IDE is not listed, gmgn-cli still works anywhere a shell does. The skills are a convenience layer; the CLI is the universal fallback.

How does the GMGN MCP-style architecture work?

People search for "gmgn mcp" expecting a Model Context Protocol server. GMGN's own description is "skill-based MCP-style architecture for agents," and the distinction is worth two sentences.

MCP servers expose tools over a protocol; the agent connects, receives a tool list with schemas, and calls tools by name. Skills achieve the same outcome with files instead of a running server: each skill folder contains a description (what the tool does, what parameters it takes) and the executable that does it. The agent reads the descriptions, picks a skill, and runs it. From the model's point of view, both look like "here is a named tool with parameters; call it when relevant."

The practical consequences:

• No daemon. Nothing runs in the background waiting for connections. Each call is a fresh process that signs a request to papi.gmgn.ai and returns JSON.
• Composable. A single natural-language request can chain several skills: trending list, then contract scoring on each result, then holder analysis on survivors, then an order.
• Inspectable. The skill code is plain and readable in the repository, so you can audit exactly what an order skill sends before you trust it with a key.
• Portable. Because it is files, the same skills work across every supported IDE without per-tool adapters.

If GMGN publishes a true MCP server later, expect it to wrap the same API and the same skill boundaries.

Example natural-language tasks

These are the kinds of prompts the skills are designed to satisfy. They map onto features that human users reach through the terminal, so the linked guides explain the underlying concepts.

• "Show me the top 20 trending Solana tokens under $2M market cap with LP burned and mint disabled." (market trending + token info; concepts in GMGN filters)
• "Score this contract for rug risk and tell me the top-10 holder share, bundler share and whether the dev has sold." (contract due-diligence + holders; see GMGN Trenches and sniper for what those signals mean)
• "Analyze wallet X: 30-day PnL, win rate, average hold time, and whether it looks like a bot." (wallet analysis; see GMGN wallet tracker)
• "Rank these five wallets for copy-trade suitability." (copy-trade wallet scoring; compare with the native feature in GMGN copy trading)
• "Pull the 15-minute K-line for this token for the last 24 hours and flag any candle with more than 3x average volume." (K-line)
• "Buy 0.1 SOL of this token with 30% slippage and 0.006 SOL priority fee, then set a 2x take-profit and 50% stop-loss." (orders; semantics in GMGN limit orders and stop-loss)
• "Show my portfolio across all wallets and total unrealized PnL." (portfolio; caveat in GMGN PnL that displayed PnL does not reliably deduct priority fees)

The first five are read-only and safe to let an agent run freely. The last two touch money and should sit behind a confirmation step.

Safety: an agent trading through GMGN is real money

Everything an order skill does, it does with your hosted GMGN wallet. There is no sandbox mode; GMGN has no paper-trading feature as of September 2026.

1. Use a dedicated account and a small balance. Create a separate GMGN account for agent use and fund it only with what a bug could lose. GMGN prohibits private key export on all chains (disabled around 19 March 2025), so the balance is a hot trading account by design, not storage. Full custody discussion in is GMGN safe.
2. Key custody. The Ed25519 private key equals trading control. Keep it in a secrets manager or a permission-restricted file, use GMGN's IP whitelist, and never include it in a prompt, a screenshot or a repo.
3. Require approval for write skills. Configure your IDE to pause before executing the order and multi-wallet skills. Language models misread numbers; "0.1 SOL" and "1 SOL" differ by one character.
4. Set fee parameters explicitly. Every trade pays 1% to GMGN plus your priority fee. GMGN recommends 0.006 SOL+ for automated orders and 30–35% slippage for automation, 50%+ on brand-new tokens. On a 0.1 SOL trade, 0.006 SOL is 6% per leg. Tell the agent your limits; do not let it guess. See GMGN fees.
5. Enable Anti-MEV where it matters. Anti-MEV needs at least 0.002 SOL priority plus 0.0001 SOL tip. GMGN launched it in July 2025 after GMGN users were found absorbing roughly 30.8% of Solana sandwich profits in an April 2025 30-day sample.
6. Handle failures as data, not retries. Codes like B4 (honeypot), D1 (liquidity pulled) and C1 (slippage revert, which burns gas) should stop the loop, not restart it.
7. Withdrawals stay manual. The API cannot withdraw; that requires the website, Google Authenticator 2FA and whitelisted addresses. Sweep profits to a self-custody wallet regularly.

What "gmgn vibe coding kit" means

"Vibe coding" is the 2025-era phrase for building software by describing it to an AI and iterating on the result rather than writing every line. The "gmgn vibe coding kit" is not a product GMGN sells; it is the community's name for the combination of gmgn-skills, a GMGN API key and an AI IDE. With those three, you can say "build me a bot that watches smart-money buys on Base, scores each token, and alerts me on Telegram when a score exceeds 80" and get a working first draft in an afternoon.

The kit's strength is speed to prototype. Its weakness is that vibe-coded trading logic is rarely tested against edge cases: partial fills, RPC outages, the "No Available Router" state during migrations (explained here), or a token that passes every check and still rugs. Prototype with it, then read the generated code before it touches a balance you care about.

Key takeaways

• gmgn skills is GMGN's official, MIT-licensed agent toolkit at github.com/GMGNAI/gmgn-skills, about 556 stars as of September 2026, with 12 skills spanning data, analysis and order placement.
• Install with npx skills add GMGNAI/gmgn-skills for AI IDEs or npm install -g gmgn-cli for a universal command-line tool; both need an API key from gmgn.ai/ai.
• It follows a skill-based MCP-style architecture: named tools with descriptions, no background server, composable in one prompt.
• Only the GMGNAI org on GitHub is official; "gmgn bot github" results asking for seed phrases are scams.
• Order skills trade a hosted wallet with real money at a 1% fee plus priority fees; use a dedicated small-balance account, gate write skills behind approval, and keep the private key out of prompts.

To set up an account for agent use, start at Open GMGN, enable 2FA, then register a key at gmgn.ai/ai.

Sources: GMGN documentation, DefiLlama, official app store listings and GMGN's public channels. See our methodology. Last reviewed 2026-09-22.