Performance-First Agent Framework That Makes Any Model Better

MiroFlow is the open-source agent framework that maximizes any model's agent performance — and proves it across 9+ benchmarks with reproducible results.
Plug in GPT-5, Claude, MiroThinker, Kimi, DeepSeek, or any OpenAI-compatible model. Same tools. Same environment. Better results.

---

📰 News

• [2026-03]: MiroFlow 1.7 + MiroThinker 1.7: Major release with Web Application interface (FastAPI + React), comprehensive verifier system for benchmark evaluation, and expanded LLM support including Kimi K2.5 and GPT-5.

Previous Updates

• [2025-09-15]: MiroFlow v0.3: Enhanced codebase architecture and significantly improved benchmark performance, boosting GPT-5's prediction accuracy for future events by 11%. MiroFlow now ranks #1 in the future prediction benchmark. See FutureX.
• [2025-08-27]: MiroFlow v0.2: Achieves state-of-the-art performance across multiple agentic benchmarks, including HLE (27.2%), HLE-Text-Only (29.5%), BrowserComp-EN (33.2%), BrowserComp-ZH (47.1%), and xBench-DeepSearch (72.0%).
• [2025-08-26]: Released GAIA Validation Trace (73.94% pass@1) and Gradio Demo for local deployment.
• [2025-08-08]: MiroFlow v0.1: Complete open-source release of the research agent framework.

---

Architecture

---

Why MiroFlow

Make Any Model Better

• Model-Agnostic Performance: Plug in any LLM — GPT-5, Claude, MiroThinker, Kimi K2.5, DeepSeek — and get better agent performance through smart rollback, iterative reasoning, and optimized tool orchestration.
• Comprehensive Benchmarking: Supports 9+ benchmarks including FutureX, GAIA, HLE, xBench-DeepSearch, BrowseComp, and more.
• One-Line Model Switching: Change provider_class and model_name in YAML. Same tools, same prompts, same environment.

Prove It

• Standardized Evaluation: Fair model comparison with identical infrastructure. The framework is the constant; the model is the variable.
• Automated Multi-Run Evaluation: Parallel runs with statistical aggregation (mean, std dev, min/max). Every result reproducible from config to score.

Build With It

• Skill System: Define agent skills via SKILL.md — no code changes needed.
• Agent Graph: Compose multi-agent workflows with hierarchical graphs.
• Web Application: FastAPI + React interface out of the box.
• Plugin Architecture: @register decorator — extend without touching core code.
• Zero-Code Prompts: YAML + Jinja2 templates.
• Cost-Effective: Single RTX 4090 with open-source MiroThinker.

---

Any Model, Better Results

Benchmark results will be updated after comprehensive testing with v1.7. See the full Model Comparison for details.

Follow our detailed guides to reproduce any result in our Benchmarks Documentation.

---

Quick Start

# 1. Clone and setup
git clone https://github.com/MiroMindAI/miroflow && cd miroflow
uv sync

# 2. Configure API keys (only OPENAI_API_KEY is required for this example)
cp .env.template .env
# Edit .env and set OPENAI_API_KEY (used by GPT-5 in the default quickstart config)

# 3. Run your first task
bash scripts/test_single_task.sh \
  --config config/agent_quickstart.yaml \
  --task-question "What is the first country listed in the XLSX file that have names starting with Co?" \
  --file-path data/FSI-2023-DOWNLOAD.xlsx

Expected output: \boxed{Congo Democratic Republic}

Switch models in one line — same tools, same prompts, different LLM:

# GPT-5
llm:
  provider_class: GPT5OpenAIClient
  model_name: gpt-5

# Claude 3.7 Sonnet
llm:
  provider_class: ClaudeAnthropicClient
  model_name: claude-3-7-sonnet-20250219

# MiroThinker (open-source, self-hosted)
llm:
  provider_class: MiroThinkerSGLangClient
  model_name: mirothinker-v1.5

See full documentation for web app setup, more examples, and configuration options.

---

References

If you find our work helpful, please consider citing:

MiroThinker (Model & Method)

@article{miromind2025mirothinker,
  title={MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling},
  author={MiroMind Team and Bai, Song and Bing, Lidong and Chen, Carson and Chen, Guanzheng and Chen, Yuntao and Chen, Zhe and Chen, Ziyi and Dong, Xuan and others},
  journal={arXiv preprint arXiv:2511.11793},
  year={2025}
}

MiroFlow (Framework)

@article{miromind2026miroflow,
  title={MiroFlow: Towards High-Performance and Robust Open-Source Agent Framework for General Deep Research Tasks},
  author={Su, Shiqian and Xing, Sen and Dong, Xuan and Zhong, Muyan and Wang, Bin and Zhu, Xizhou and Chen, Yuntao and Wang, Wenhai and Deng, Yue and Zhu, Pengxiang and others},
  journal={arXiv preprint arXiv:2602.22808},
  year={2026}
}

---

Contributing: Issues · Pull Requests · Discord

License: Apache 2.0