MiroFlow is a fully open-source research agent framework built for multi-step internet research on hard problems: forecasting future events, answering questions that require browsing and cross-checking many sources, and completing complex tasks that single prompts cannot handle.
MiroFlow is the official implementation of the MiroMind Research Agent Project, developed on GitHub, where it has collected more than 3,100 stars under the Apache-2.0 license, with active development through 2026. The team reports top-1 results on five or more representative benchmarks, including FutureX, GAIA, HLE, xBench-DeepSearch, and BrowseComp, with reproducible configurations so others can verify the scores. The framework pairs with MiroThinker, the project’s open agent model with native tool-assisted reasoning, but it is deliberately model-agnostic: it also supports Claude, Kimi, and OpenAI models, so you can benchmark the same research workflow across providers. A web UI is included for running and watching research sessions without living in the terminal, and the project reports that its v0.3 release boosted GPT-5’s future-prediction accuracy on FutureX by eleven percent. MiroFlow is part of a broader open ecosystem that includes MiroVerse, a large open dataset for training research agents, making the whole stack, data, model, and framework, available for inspection and extension.
MiroFlow is free and open source under the Apache-2.0 license. You run it yourself and pay only for the model API usage your research tasks consume.
MiroFlow is for agent researchers, teams building deep-research products, and benchmark engineers who need a strong, transparent baseline for web research agents instead of a closed API they cannot inspect or reproduce.