wtf is gmgn-minicpm?

gmgn-minicpm lets a small AI model run entirely on your own computer to answer crypto research questions using the GMGN data platform, instead of sending your questions to a cloud AI provider like Claude or GPT. It uses MiniCPM5-1B, a roughly one billion parameter model small enough to run locally, and connects it to the existing gmgn-cli command line tool through a small TypeScript bridge with no other dependencies. When you ask a question such as what smart money is buying on Solana right now, the local model decides which gmgn-cli command answers it, the bridge runs that command as a plain subprocess without a shell, and the resulting JSON data is fed back to the model so it can form an answer. Because the model and the bridge both run on your own machine, there is no ongoing API bill for the AI itself. You only pay for GMGN's own data access. This is explicitly a read-only, version one release. Only research tools that look up token, market, wallet, and portfolio information are registered with the model. Any command that could move funds, such as swap or order commands, is left out of the tool list entirely and is also actively blocked inside the bridge, since a small model running in an automated loop should never be able to move money on its own. The documentation is clear that this tool only reads data and provides no financial advice. Running it requires a recent version of Node.js, the gmgn-cli tool configured with an API key, and a local server exposing MiniCPM5-1B through an OpenAI style chat interface, either llama.cpp on Apple Silicon or CPU, or SGLang on NVIDIA hardware. The project also includes safeguards for working with a small model, such as capping how many tokens it can generate and truncating oversized tool results, since a 1B model can otherwise spiral into invalid output. Planned future work includes fine-tuning the model on GMGN specific examples and adding fund-moving actions behind a mandatory human confirmation step. The project is released under the MIT license.