gmgn-minicpm runs a 1B model on your own machine as a read-only GMGN crypto research agent: no cloud LLM, no API cost, offline-capable. Most people wiring GMGN into an agent point it at Claude or GPT; this project points it at MiniCPM5-1B, a ~1B on-device model with native tool calling, running locally. The project itself is a thin, dependency-free TypeScript bridge that turns the model's tool_calls into safe gmgn-cli commands and feeds the JSON back.

Plain Text

  "what's smart money buying on sol right now?"
                 │
                 ▼
        MiniCPM5-1B  ──(OpenAI-style tool_calls)──►  gmgn-local bridge
        (local, SGLang)                                   │
                 ▲                                        ▼  (execFile, no shell)
                 └────────────(JSON result)──────  gmgn-cli track smartmoney --chain sol --raw
                                                          │
                                                          ▼
                                                     GMGN OpenAPI

Why

• No API bill: the model runs on your hardware; you only pay GMGN's data quota.
• Low latency / offline: no cloud round trip, which suits always-on watchers.
• Small and auditable: zero runtime dependencies, native Node TypeScript, and every tool is an explicit allowlisted subcommand.

Safety

This is v1: read-only. Only market, token and wallet research tools are exposed. Fund-moving commands (swap, multi-swap, order, cooking) are not registered and are also hard-blocked in the bridge. A 1B model in an automated loop must never be handed the ability to move funds. If swap support lands, it will sit behind explicit human confirmation.

Nothing here is financial advice. It reads data; you make the decisions.

How It Works

• src/tools.ts: the read-only tool registry. Each entry is an OpenAI-style function schema plus the gmgn-cli subcommand it maps to; snake_case params become --kebab-case flags.
• src/bridge.ts: validates a tool call (required args, drops hallucinated args, blocks fund-moving roots) and builds an argv array.
• src/gmgnCli.ts: runs gmgn-cli with execFile (argv array, no shell, so args cannot be injected), appends --raw, and parses the JSON.
• src/agent.ts: the model loop — send messages and tools → run returned tool_calls → feed results back → repeat until a final answer (bounded by GMGN_LOCAL_MAX_ITERS).

Tools (v1)

• token: info · security · pool · holders · traders
• market: kline · trending · trenches · signal · hot-searches
• portfolio: info · holdings · activity · stats · token-balance · created-tokens
• track: follow-tokens · follow-wallet · kol · smartmoney

Running a 1B Model Reliably

A 1B model is not a frontier model, and the repo is shaped around that:

• Cap generation (GMGN_LOCAL_MAX_TOKENS, default 1024). Small models can spiral on an error and emit runaway tool-call JSON until the context fills, which makes llama.cpp return a hard 500. Capping tokens prevents it.
• Truncate tool results (GMGN_LOCAL_MAX_TOOL_RESULT_CHARS, default 4000). A single trending response is ~7 KB; a few unbounded results overflow an 8k context. Serve with a roomy -c 32768.
• Validate enums in the bridge. The model occasionally invents a chain like "5"; bridge.ts rejects it with must be one of: sol, bsc, base, eth so the model can self-correct instead of shelling out garbage.

These are exactly the rough edges a LoRA fine-tune is meant to smooth out.

Quick Start

Requirements:

• Node.js ≥ 24 (runs the TypeScript sources directly via native type stripping, no build step).
• gmgn-cli installed and configured (npm install -g gmgn-cli, then gmgn-cli config).
• A local MiniCPM5-1B server exposing an OpenAI-compatible /v1/chat/completions with tool calls:
  • Apple Silicon / CPU: llama.cpp llama-server is the verified path; with --jinja it parses MiniCPM5's tool calls into native OpenAI tool_calls.
  • NVIDIA / CUDA: SGLang with --tool-call-parser minicpm5.

Verified: Apple M4 Pro / 48 GB, MiniCPM5-1B F16 GGUF on llama-server, Metal-accelerated at ~90–100 tok/s. The full loop (model → gmgn-cli → answer) has been confirmed on live GMGN data.

Bash

git clone https://github.com/dvictor357/gmgn-minicpm
cd gmgn-minicpm
cp .env.example .env

# 1. serve the model (auto-downloads the F16 GGUF from HF on first run)
bash scripts/serve-llamacpp.sh          # NVIDIA users: scripts/serve-sglang.sh

# 2. ask something
node src/cli.ts "what is smart money buying on solana right now?"

# or an interactive REPL
node src/cli.ts

Add --verbose to print each tool call the model makes; node src/cli.ts --list-tools lists the tools without a model running.

Roadmap

• LoRA dataset generator: npm run dataset builds a (query → tool_call → gold answer) set from real GMGN data with deterministic gold labels.
• LoRA fine-tune and fuse to GGUF (MLX-LM on Apple Silicon) for reliable composite routing, argument selection and concise judgment.
• Gated swap: fund-moving tools behind a mandatory human-confirmation / rule-engine gate, never a raw model decision.
• 24/7 watcher mode (monitor trending/signals → alert).

Links

• GitHub: dvictor357/gmgn-minicpm