持续观察 · 2026-10-16Continuous watch · 2026-10-16

热门,不等于
值得投入。
Trending is not
the same as useful.

每天记录项目变化,在重要事件发生时重新分析。我们把 GitHub 动量、实际用途、成熟度、成本与风险分开评分,让你知道一个项目为什么火,也知道要不要花时间。

We capture project changes daily and reassess when meaningful events occur. Momentum, utility, maturity, cost and risk stay separate—so you can tell why a project is hot and whether it deserves your time.

12 个观察项目12 projects watched每日 GitHub 快照Daily GitHub snapshots评分可追溯Traceable scoring无付费排名No paid rankings
ACTIVE WATCHLIST

当前观察名单Active watchlist

指标来自 GitHub API;评分是编辑判断。新增项目先做官方资料评审,经过安装实测和连续快照后再提高置信度。

Metrics come from the GitHub API; scores are editorial. New entries begin with an official-source review, then gain confidence through hands-on testing and longitudinal data.

01 / AGENT TOOL78

REA

morluto/rea

让 Agent 从应用行为一路分析到原生二进制。势头强、范围大,但高权限分析和结果复核成本不能忽略。

Agents reverse-engineer targets from app behavior to native binaries. High momentum and broad scope, with non-trivial privilege and verification costs.

432,121Stars
107,373Forks
CONFIDENCE · 61%WATCH
02 / PERSONAL AGENT82

OpenClaw

openclaw/openclaw

运行在个人设备和聊天渠道里的开源助手。生态、活跃度和跨平台能力突出,安全边界与长期维护是关键。

An open assistant across personal devices and chat channels. Strong ecosystem and reach; security boundaries and operational upkeep remain decisive.

393,935Stars
82,785Forks
CONFIDENCE · 68%LEADER
03 / DECISION MODEL57

Jev

jevmodel/jev-ai

将状态转成带概率的类型化决策。新闻话题明显,但公开仓库只是产品说明,代码活跃度不能验证产品成熟度。

Turns state into typed decisions with probabilities. Newsworthy concept; the public repository is a product page, not evidence of implementation maturity.

0Public Stars
0Public Forks
CONFIDENCE · 42%EARLY
04 / CODING AGENT84

Gemini CLI

google-gemini/gemini-cli

Google 的开源终端 Agent,覆盖代码理解、文件修改、Shell、MCP 与自动化。能力完整,但模型依赖和执行权限需要认真配置。

Google's open terminal agent for code, files, shell, MCP and automation. Broad capability, with model dependency and execution permissions to manage.

108,732Stars
14,917Forks
CONFIDENCE · 72%STRONG
05 / BROWSER AGENT79

Browser Use

browser-use/browser-use

让 AI 操作真实浏览器的 Python 库,并提供托管 Agent 与浏览器基础设施。价值直接,权限与云端成本边界必须分清。

A Python library for agents that operate real browsers, plus hosted agents and browser infrastructure. Useful, but permissions and cloud costs need separation.

120,074Stars
13,394Forks
CONFIDENCE · 65%WATCH
06 / DOCUMENT AI76

MarkItDown

microsoft/markitdown

把 PDF、Office、网页和其他文件转成适合 LLM 的 Markdown。用途清楚、接入简单,转换保真与不可信文件风险仍需实测。

Converts PDFs, Office files, web pages and more into LLM-friendly Markdown. Clear utility; fidelity and untrusted-file handling still need testing.

193,947Stars
14,569Forks
CONFIDENCE · 70%USEFUL
07 / CODING AGENT85

Claude Code

anthropics/claude-code

Anthropic 的终端编程 Agent,可处理代码库、Git 工作流和自动化。产品成熟,但公开仓库不采用标准开源许可证。

Anthropic's terminal coding agent for codebases, Git workflows and automation. Mature product; the public repository is not under a standard open-source license.

154,586Stars
25,566Forks
CONFIDENCE · 72%STRONG
08 / CODING AGENT86

Codex CLI

openai/codex

OpenAI 的本地终端编程 Agent,支持 ChatGPT 登录或 API Key,并与 IDE、桌面和云端产品形成完整工作流。

OpenAI's local terminal coding agent with ChatGPT or API authentication, connected to broader IDE, desktop and cloud workflows.

133,262Stars
21,107Forks
CONFIDENCE · 75%LEADER
09 / CODING AGENT82

Qwen Code

QwenLM/qwen-code

面向终端、IDE、桌面和聊天渠道的开源编程 Agent,支持 Qwen、第三方 API 与本地模型,选择自由度高。

An open coding agent across terminal, IDE, desktop and chat, with Qwen, third-party API and local-model support.

29,016Stars
3,315Forks
CONFIDENCE · 66%WATCH
10 / CODING AGENT85

OpenCode

anomalyco/opencode

模型供应商中立的开源编程 Agent,终端与桌面体验成熟、社区增长强,但权限确认不等于安全隔离。

A provider-neutral open coding agent with mature terminal and desktop experiences. Strong momentum; permission prompts are not isolation.

224,667Stars
30,279Forks
CONFIDENCE · 76%LEADER
11 / CODING AGENT84

Cline

cline/cline

横跨 IDE、CLI 与 SDK 的开源编程 Agent,模型选择和工具链完整;Auto Approve、MCP 与本地执行权限需要谨慎配置。

An open coding agent across IDE, CLI and SDK. Broad model and tool support, with Auto Approve, MCP and local execution permissions to configure carefully.

71,331Stars
7,863Forks
CONFIDENCE · 78%STRONG
12 / AGENT HARNESS80

DeepSeek Harness

deepseek-ai/deepseek-harness

以“一切皆插件”为核心的开源 Agent Harness,组合能力和增长惊人;仍处于开发者预览,兼容性与安全边界必须先验证。

An everything-is-a-plugin agent harness with striking composability and momentum. Still a developer preview, so compatibility and safety need validation first.

273,104Stars
33,283Forks
CONFIDENCE · 60%EARLY

英文项目指南:Full guides in English: REA · OpenCode. Other cards have English summaries and Chinese full reviews.

先把权限边界设对。Set the security boundary first. 审批不是沙箱,Checkpoint 也不能撤销外部操作:Approvals are not a sandbox, and checkpoints cannot undo external side effects: 阅读 Agent 安全使用指南 →Read the agent safety guide →
想按工作流和许可筛选工具:Need to filter tools by workflow and license: 使用工具选择器 →Use the tool chooser →
Roo Code 已停止服务:Roo Code has shut down: 查看迁移方案 →See migration options →
本机实测:Mac mini M4/32GB 跑 Qwen。Hands-on: Qwen on a Mac mini M4 / 32GB. 9B、27B、35B 的生成速度、首字等待、内存口径和专家卸载代价,附脱敏逐次数据。Decode speed, first-token latency, memory units and expert-offload tradeoffs, with sanitized per-run data. 阅读实测与限制 →Read the field report (Chinese) →
HOW WE READ SIGNALS

把事实和判断拆开。Facts and judgment stay separate.

01 · SNAPSHOT

数据事实Observed metrics

Star、Fork、Release、Commit 与 Issue 来自公开接口,并保存采集时间。没有历史就不显示伪造的增长率。

Stars, forks, releases, commits and issues come from public interfaces with capture times. We do not invent growth rates without history.

02 · REVIEW

编辑判断Editorial judgment

用途、成熟度、安全、成本分别评分。总分不是客观真理,每个分数都应能追溯到理由。

Utility, maturity, safety and cost are scored separately. The total is an editorial judgment, and each score needs a traceable reason.

03 · CHANGE

变化解释Meaningful changes

数据每日记录;只有版本、异常增长、安全事件或方向变化时,才重写结论和评分。

We record metrics daily and revise conclusions when releases, unusual growth, security events or project direction change.

当前限制:新增项目暂基于官方仓库、文档与公开指标评审,尚未完成统一环境安装实测。页面会明确区分“已验证”“官方声明”和“编辑推断”,评分置信度也会随证据更新。Current limit: New projects are initially reviewed from official repositories, documentation and public metrics. We have not installed every project in one controlled environment. Pages distinguish verified findings, vendor claims and editorial inference; confidence changes as evidence improves.