MOMO CODE
An AI coding agent, built on opencode, that adjusts its own tactics from session outcomes through a dual-speed self-evolution system, with recursive subagents and a long-running task engine.
MOMO CODE is an AI coding agent built on opencode that adds a two-speed self-evolution system layered on top of it: a fast experience loop selects tactics distilled from past successes through Thompson sampling, while a slower self-evolution training mode improves weights over an hour-scale run using Monte Carlo Graph Search and LoRA. A self-refinement command reviews session trajectories and proposes small, evidence-based prompt or tactic patches that only take effect after a human approves them.
It supports more than 25 LLM providers, including DeepSeek, Zhipu, Moonshot, Claude, GPT-4 and Gemini, plus custom OpenAI-compatible endpoints, and adds recursive subagent task decomposition, a resumable dependency-graph engine for long-horizon tasks, a persistent-goal daemon loop, and voice input through local speech-to-text. It can inherit an existing .claude/ configuration, including MCP servers and prompts, for migration from Claude Code.