# MiroThinker — Open Deep Research Agent with 256K Context

> MiroThinker is an open deep-research agent stack with long context, tool-heavy configs, and benchmarked BrowseComp results for serious evaluation work.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

1. Clone and set up the agent app:
   ```bash
   git clone https://github.com/MiroMindAI/MiroThinker
   cd MiroThinker/apps/miroflow-agent
   uv sync
   cp .env.example .env
   ```
2. Fill in the required keys such as `SERPER_API_KEY`, `JINA_API_KEY`, and `E2B_API_KEY`.
3. Verify:
   - Load one recommended config like `mirothinker_1.7_keep5_max200` and complete a single research task.

## Intro

MiroThinker is an open deep-research agent stack with long context, tool-heavy configs, and benchmarked BrowseComp results for serious evaluation work.

- **Best for:** research teams benchmarking long-horizon, tool-rich open agents
- **Works with:** Python 3.10+, uv, E2B sandbox, Serper, Jina, summary LLMs, YAML agent configs
- **Setup time:** 30-60 minutes

## Practical Notes

- Quant: MiroThinker-1.7 is documented with 256K context and up to 300 tool calls per task; older v1.0 notes mention up to 600 calls.
- Quant: the README highlights BrowseComp 74.0 and BrowseComp-ZH 75.3 for the 1.7 line, plus multiple keyed environment variables for the minimal tool set.

## Why it matters

MiroThinker is better viewed as a research reproduction stack than a plug-and-play assistant, and that is exactly why it is valuable for serious agent benchmarking.

- The README is unusually concrete about context limits, tool counts, and benchmark scores, which gives evaluators something measurable to reason about.
- The minimal configuration section explains the exact tool servers needed for search, scraping, and code execution.
- Pre-configured agent YAMLs make it easier to compare runs without inventing a fresh orchestration setup every time.

## Rollout pattern

- Treat the minimal tool set as the baseline and resist adding optional tools until you can reproduce one benchmark-like task.
- Log environment variables and configs per run so your benchmark notes are reproducible.
- Use it to learn what long-horizon research agents require operationally, not as a trivial one-command chatbot replacement.

## Watchouts

This stack has more moving parts than a typical app-facing agent, so missing keys or mismatched tool configs will produce noisy failures unless you document the environment carefully.

### FAQ

**Q: Is it a simple local chat app?**
A: No. The README positions it as a tool-enabled research agent with multiple required services and configs.

**Q: Why is it worth studying?**
A: Because it publishes measurable context, tool-call, and benchmark facts that are useful for replication.

**Q: What should I verify first?**
A: One recommended YAML config plus the three-tool minimal setup for search, scraping, and execution.

## Source & Thanks

> Source: https://github.com/MiroMindAI/MiroThinker
> License: Apache-2.0
> GitHub stars: 8,223 · forks: 626

---

<!-- ZH -->

## 快速使用

1. 克隆并设置 agent 应用：
   ```bash
   git clone https://github.com/MiroMindAI/MiroThinker
   cd MiroThinker/apps/miroflow-agent
   uv sync
   cp .env.example .env
   ```
2. 补齐 `SERPER_API_KEY`、`JINA_API_KEY`、`E2B_API_KEY` 等必需密钥。
3. 验证：
   - 载入推荐配置 `mirothinker_1.7_keep5_max200`，完成一次研究任务。

## 简介

MiroThinker 是一套面向深度研究场景的开源 Agent 栈，强调长上下文、密集工具调用和基准成绩，适合希望复现实验、研究配置并系统评估研究型 Agent 能力边界与操作成本的团队。

- **适合谁：** 要评估长链路、重工具开源研究 Agent 的研究团队
- **可搭配：** Python 3.10+、uv、E2B 沙箱、Serper、Jina、摘要模型与 YAML agent 配置
- **准备时间：** 30-60 分钟

## 实战建议

- 量化信息：README 为 MiroThinker-1.7 标注了 256K 上下文与单任务最多 300 次工具调用；更早版本甚至写到 600 次。
- 量化信息：README 同时给出 BrowseComp 74.0、BrowseComp-ZH 75.3 等成绩，并列出了最小工具集所需的多项环境变量。 

## 为什么值得收录

MiroThinker 更像“研究复现实验栈”而不是即装即用聊天助手，而这恰恰是它的价值所在。

- README 对上下文长度、工具调用次数和基准分数给得很具体，方便评估者做可量化判断。
- 最小配置段把搜索、抓取和执行所需的工具服务器写得很清楚。
- 预配置 YAML 能帮助你在不同实验之间复用编排，而不是每次重新搭框架。

## 落地路径

- 先严格按最小工具集复现，不要一开始就叠加可选工具。
- 每次运行都记录环境变量与配置文件，保证实验笔记可复验。
- 把它当作研究型 Agent 的操作学习样本，而不是轻量聊天工具替代品。 

## 注意事项

它的活动部件明显多于普通 Agent，若不把密钥和配置记录清楚，失败往往会变得噪音很大、难以复盘。

### FAQ

**它是简单的本地聊天应用吗？**
答：不是。README 把它定义为需要多种服务与配置的工具增强型研究 Agent。

**为什么值得研究？**
答：因为它公开了上下文、工具调用和 benchmark 这些可量化事实，适合复现实验。

**第一步该验证什么？**
答：先验证 1 个推荐 YAML 配置和最小三工具组合。

## 来源与感谢

> Source: https://github.com/MiroMindAI/MiroThinker
> License: Apache-2.0
> GitHub stars: 8,223 · forks: 626


---
Source: https://tokrepo.com/en/workflows/mirothinker-open-deep-research-agent-with-256k-context
Author: Agent Toolkit