跳到主要内容Skip to main content
投机取巧 Speculative Speculation

AI 工具AI Tools · Tools

给 AI 装工具Tools for your AI

MCP server 的权限范围、传输方式与输出可用性。收录标准：已发布为可安装的 MCP server、有官方仓库或文档可回溯、近 30 天有实质更新。这里是能力缺口查表，不是选型对比。The permission scope, transport and output usability of MCP servers. Admission criteria: released as an installable MCP server, has an official repository or documentation that can be traced back, and has substantive updates within the last 30 days. This is a capability-gap reference table, not a selection comparison.

10 个对象objects 11 个固定维度fixed dimensions 8 个已完整核验fully verified 最后核验Last checked 2026-10-08
  1. 这里收录能接任务的产品。它要么在编辑器或终端里替你写代码，要么能被你装进自己的程序，要么能连上你的文件和账号。This site lists products that take on tasks. They either write code for you in your editor or terminal, or can be installed into your own program, or connect to your files and accounts.
  2. 纯聊天工具不在这里。如果你只想找一个对话问答的网页，去 AI 导航，ChatGPT、豆包、Kimi 这些都在那里。Plain chat products are not here. If you just want a chat webpage, go to the AI Directory — ChatGPT, Doubao and Kimi are listed there.
  3. 每条档案都是核验快照。字段逐条挂官方出处与核验日期，本站不实测、不排名，未知就写未知。Every entry is a verification snapshot. Each field carries an official source and a checked date; we run no benchmarks, publish no rankings, and write "unknown" when it is unknown.

按任务找Find by task

这一页有 10 个对象。下面的数字是本页真正收录的条数——点进去就是筛选后的结果，不是跨站的总数。This page holds 10 objects. The numbers below are what this page actually carries — clicking filters right here, not a site-wide total.

陪你创作Create with AI

这些工具不走八维坐标系——它们没有「安装形态 / 工具清单」这些维度，硬套只会产出一堆「不适用」。创作工具真正可比的是「能做什么 / 数据去哪 / 适合谁 / 输出物能不能商用」；商用许可逐条抄官方条款（许可稳定，值得存），价格常变，本站不存报价，一律看官方定价页。These tools are not on the 8-axis grid — they lack dimensions like install form or tool inventory, and forcing them would produce a wall of "not applicable". What is genuinely comparable for creation tools is what they do, where your inputs go, who they suit, and whether the output may be used commercially. Commercial licensing is quoted line by line from the official terms (licences are stable, worth storing); prices change often, so this site stores no quotes and always points to the official pricing page.

MidjourneyMidjourney

Midjourney, Inc. · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 直连未核验direct unknown 未见于备案名单not in filing list

能做什么What it does
图像与视频生成；网页应用与 Discord Bot 两条使用线；社区公开浏览与 Remix 是产品默认形态。Image and video generation via the web app and the Discord bot; public browsing and remixing are the product's default community shape.
数据去哪Where your data goes
官方隐私政策 §1 明确其适用范围包含「通过训练 Midjourney 机器学习算法的过程收集的数据」；官方 AB2013 训练数据披露称训练集含「第三方与 Midjourney 用户提供的数据」。官方帮助页称默认公开可浏览、可被他人 Remix，仅 Pro / Mega 档的 Stealth 模式承诺尽力不公开。官方文档未见面向用户的训练退出开关。The official privacy policy §1 states it applies to data 'collected through the process of training Midjourney machine learning algorithms'; the official AB2013 training-data disclosure lists 'data our users provide through use of the Midjourney service' among training sources. Official docs describe content as publicly viewable and remixable by default, with Stealth Mode (Pro / Mega only) as the best-efforts non-publishing option. No user-facing training opt-out control is documented.
适合谁Who it suits
重图像风格质量、能接受「默认公开」社区形态的人。要作品不公开或商用归属清晰，先看订阅档与 Stealth 条款再开工。For people who weight image style quality and accept an open-by-default community. If non-public work or clean commercial ownership matters, read the plan tiers and Stealth terms before starting.
商用许可Commercial use
官方 ToS §4「Content Rights」：付费用户拥有 Assets——「You own all Assets You create with the Services to the fullest extent possible under applicable law」，且「Your ownership of the Assets you created persists even if in subsequent months You downgrade or cancel Your membership」；例外：年营收超 100 万美元的公司（或其员工）必须订阅 Pro / Mega 档才拥有 Assets；他人图像 upscale 后仍归原作者。非付费用户不拥有 Assets，只获 CC BY-NC 4.0（非商用）许可。另授予 Midjourney 对输入与 Assets 的永久免许可金。Official ToS §4 'Content Rights': paid users own their Assets — 'You own all Assets You create with the Services to the fullest extent possible under applicable law', and ownership 'persists even if in subsequent months You downgrade or cancel Your membership'. Exceptions: companies (or their employees) with more than $1,000,000 USD annual revenue must subscribe to a Pro or Mega plan to own Assets; upscaling another user's image leaves it owned by the original creator. Non-paid members do not own Assets — they receive a Creative Commons Noncommercial 4.0 license. Users also grant Midjourney a perpetual royalty-free license over Content and Assets.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：2 条官方链接 · 核验 2026-10-20 · verifiedSources: 2 official links · checked 2026-10-20 · verified

Stability AI（Stable Image 平台）Stability AI (Stable Image platform)

Stability AI · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 大陆可直连direct OK 未见于备案名单not in filing list

能做什么What it does
图像生成平台（Stable Image 系列模型）；另有可自托管的开源权重模型（Stable Diffusion 系）。An image-generation platform built on the Stable Image models; separately, open-weight Stable Diffusion models can be self-hosted.
数据去哪Where your data goes
官方 ToS：用户可选择退出把内容用于训练（opt-out 机制在条款内写明）；平台对上传内容取得运营所需许可。开源权重模型的训练数据口径见官方模型卡与基准报告。Per the official Terms of Service, users may opt out of having their content used for training (the opt-out mechanism is written into the terms); the platform takes the licenses it needs to operate. Training-data disclosures for the open-weight models live in the official model cards and benchmark reports.
适合谁Who it suits
想要「平台服务 + 可自托管开源模型」两条线都摸得到的人。商用条款平台版与开源版不同，分开读。For people who want both a hosted platform and self-hostable open weights. Platform licensing and open-model licensing are different documents — read them separately.
商用许可Commercial use
官方 ToS §4(a)：「we assign to you all of our right, title, and interest (if any) in the Outputs」——输出物所有权归用户；§3(b) 限制：不得转售输出、不得用输出训练竞争模型、不得去除水印或来源标识。开源的 Stable Diffusion 权重按各自的开源许可证发布（自托管商用看模型许可证）。Official ToS §4(a): 'we assign to you all of our right, title, and interest (if any) in the Outputs' — output ownership is assigned to the user. §3(b) limits: no reselling outputs, no training competing models on outputs, no removing watermarks or provenance marks. The open-weight Stable Diffusion models are released under their own open licenses; commercial self-hosting follows those model licenses.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：1 条官方链接 · 核验 2026-10-20 · verifiedSources: 1 official links · checked 2026-10-20 · verified

FLUX（Black Forest Labs）FLUX (Black Forest Labs)

Black Forest Labs · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 大陆可直连direct OK 未见于备案名单not in filing list

能做什么What it does
图像生成模型家族（FLUX 系）：官方 API 与 playground 提供托管生成，部分权重开源可自托管。The FLUX family of image models: hosted generation through the official API and playground, with some weights open for self-hosting.
数据去哪Where your data goes
官方 ToS：默认可能用输入与输出改进模型，退出走 [email protected] 邮件申请；条款对「用输出训练其他模型」有明确限制。Per the official ToS, inputs and outputs may be used to improve the models by default; opt-out is requested by emailing [email protected]. The terms explicitly restrict training other models on outputs.
适合谁Who it suits
要接 FLUX 官方 API、或想自托管开源 FLUX 权重的人。托管与自托管的许可路径不同，商用前先确认走哪条。For people using the official FLUX API or self-hosting open FLUX weights. Hosted and self-hosted licensing are different paths — confirm which one you are on before commercial use.
商用许可Commercial use
官方 ToS 1.2(a)：平台不主张输出物所有权；1.3(e)：商用权限于用户自己的 Output；1.3(n)：禁止用 Output 训练其他模型。开源 FLUX 权重按各自的开源许可证发布（自托管商用看模型许可证）。Official ToS 1.2(a): the platform does not claim ownership of Outputs. 1.3(e): commercial use is limited to your own Outputs. 1.3(n): training other models on Outputs is prohibited. Open FLUX weights are released under their respective open licenses; commercial self-hosting follows those model licenses.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：1 条官方链接 · 核验 2026-10-20 · verifiedSources: 1 official links · checked 2026-10-20 · verified

Veo（Google）Veo (Google)

Google · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 大陆可直连direct OK 未见于备案名单not in filing list

能做什么What it does
Google 的视频生成模型家族：经 Gemini API / Vertex AI 与 Google AI 订阅计划提供，带水印与 SynthID 溯源标记。Google's video-generation model family, available through the Gemini API / Vertex AI and Google AI plans, with watermarks and SynthID provenance marks.
数据去哪Where your data goes
官方 Gemini API 附加条款（2026-03-23 生效）：非付费服务（AI Studio、免费配额）提交的内容与生成回答会被用于「提供、改进与开发 Google 产品与机器学习技术」，且人工审核员可能阅读；付费服务不用于改进产品，按数据处理附加条款处理。EEA / 瑞士 / 英国用户即使免费配额也按付费口径。Per the official Gemini API Additional Terms (effective 2026-03-23): for Unpaid Services (AI Studio, free quota), submitted content and generated responses are used 'to provide, improve, and develop Google products and services and machine learning technologies', and human reviewers may read them; for Paid Services, prompts and responses are not used to improve products and are handled under the Data Processing Addendum. Users in the EEA / Switzerland / the UK get the Paid-Services data terms even on free quota.
适合谁Who it suits
已在 Google 云 / Workspace 生态、需要企业级数据条款（付费不训练）的视频生成用户。免费配额线会被训练，别混。For video generation inside the Google cloud ecosystem that needs enterprise data terms (no training on paid tiers). The free-quota line is trained on — don't conflate the two.
商用许可Commercial use
官方 Gemini API 附加条款「Use of Generated Content」：「Google won't claim ownership over that content」——Google 不主张生成内容所有权；商用按适用法律合规使用（API 返回时可能需向你的用户提供归属信息）。官方同时声明输出可能与他人相似、并保留水印与 SynthID 标记。Official Gemini API Additional Terms, 'Use of Generated Content': 'Google won't claim ownership over that content.' Commercial use must comply with applicable law, and content returned via API calls may require attribution to your users. Google also notes outputs may be similar across users and reserves watermarking and SynthID marks.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：2 条官方链接 · 核验 2026-10-20 · verifiedSources: 2 official links · checked 2026-10-20 · verified

RunwayRunway

Runway AI, Inc. · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 大陆可直连direct OK 未见于备案名单not in filing list

能做什么What it does
视频与图像生成平台（Gen 系模型 + API）；API 可嵌入自有应用（须按品牌指引标注 Powered by Runway）。A video and image generation platform (the Gen model family plus APIs); APIs can be embedded in your own applications subject to 'Powered by Runway' branding guidelines.
数据去哪Where your data goes
官方 ToU §4.4：输入与输出可能被用于训练与改进其模型——条款写明用户授予「non-exclusive, irrevocable, perpetual」训练许可；§4.2 平台不主张 Your Content 所有权，但公开区内容会授予大范围许可。未见训练退出开关。Per the official Terms of Use §4.4, Inputs and Outputs may be used to train and improve Runway's models — the clause grants a 'non-exclusive, irrevocable, perpetual' license for that purpose. §4.2 states the company does not claim ownership of Your Content, while publicly posted content carries a broad license. No training opt-out control is documented.
适合谁Who it suits
重视频生成工作流、要 API 集成的团队。在意「素材绝不进训练」的要先确认这一点——条款层面没有退出选项。For teams weighting video-generation workflows and API integration. If 'my material never enters training' is a hard requirement, note the terms offer no opt-out.
商用许可Commercial use
官方 ToU §4.4「User Inputs and Outputs」：「The Company does not claim ownership of any of your Inputs or Outputs. Subject to your compliance with the Agreement, the Company does not restrict your commercial use of your Outputs.」——不主张所有权、商用不设限（以遵守条款为前提）；§10.1(a) 声明输出可能不唯一。§5(viii) 禁止用输出开发竞争产品。Official Terms of Use §4.4 'User Inputs and Outputs': 'The Company does not claim ownership of any of your Inputs or Outputs. Subject to your compliance with the Agreement, the Company does not restrict your commercial use of your Outputs.' §10.1(a) warns outputs may not be unique. §5(viii) bars using outputs to develop competing products.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：1 条官方链接 · 核验 2026-10-20 · verifiedSources: 1 official links · checked 2026-10-20 · verified

可灵 Kling AIKling AI

Kling AI Pte. Ltd.（快手） · 轻档案 · 不上八维light entry · not on the 8-axis grid

国产domestic 大陆可直连direct OK 已备案filed

能做什么What it does
视频与图像生成（Kling 视频系 + Image 系）；网页 / App / API 三线，另有 Team Space 团队空间。Video and image generation (the Kling video and image model families) across web, app and API, with Team Spaces for collaboration.
数据去哪Where your data goes
官方国际版 ToS（2026-04-21 生效）§4.7.3(f)：平台可为「create, test, improve, train」其模型处理内容；§4.7.4 退出走邮件 [email protected]（撤销面向未来，不溯及已完成的训练）。§4.8：发布到社区的内容可被其他用户 Recreate / 下载。Per the official international ToS (effective 2026-04-21) §4.7.3(f), Kling may process content to 'create, test, improve, train' its models; §4.7.4 offers revocation by emailing [email protected] — forward-looking only, not unwinding completed training. §4.8: content published to the community may be recreated or downloaded by other users.
适合谁Who it suits
重图生视频质量与中文生态的人。商用要先过会员线（见许可列），发布到社区等于公开。For people weighting image-to-video quality and the Chinese-language ecosystem. Commercial use runs through the membership line (see the license column); publishing to the community means publishing.
商用许可Commercial use
官方国际版 ToS §4.4：「You own all intellectual property rights and other proprietary interests in and to the Content under applicable law… We do not claim ownership of the Content.」——所有权归用户；但 §4.6：「without our written permission, you may not use, reproduce, distribute… the Output for any commercial purposes」——商用需书面许可。会员的商业许可在《Terms of Paid Service》（3.1.2 会员商用不受限、7.2 水印归属）——该页正文为动态加载、本次直抓未能取得，本站不转述其细节；非会员按 §4.5 须为输出标注 Kling 品牌。Official international ToS §4.4: 'You own all intellectual property rights and other proprietary interests in and to the Content under applicable law… We do not claim ownership of the Content.' But §4.6: 'without our written permission, you may not use, reproduce, distribute… the Output for any commercial purposes' — commercial use requires written permission. The membership commercial license lives in the Terms of Paid Service (reported as 3.1.2 members' commercial use unrestricted and 7.2 watermark rights) — that page loads its text dynamically and could not be fetched directly on 2026-10-20, so this site does not restate its details. Non-members must label output with Kling branding per §4.5.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：2 条官方链接 · 核验 2026-10-20 · partialSources: 2 official links · checked 2026-10-20 · partial

SunoSuno

Suno Inc. · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 直连不可达direct blocked 未见于备案名单not in filing list

能做什么What it does
AI 音乐生成（歌曲 + 人声 + Voice Model 声音克隆）；网页与移动端。AI music generation — songs, vocals and Voice Models (voice cloning) on web and mobile.
数据去哪Where your data goes
官方 ToS：用户保留 Submissions 所有权（「you retain all right, title, and interest in and to your Submissions」），但授予 Suno 全球免许可金（含创建衍生作品）；条款明确该许可覆盖「train, develop, fine-tune or otherwise improve」其模型，未见训练退出开关。禁止用输出训练其他模型。Per the official ToS, users retain all right, title and interest in their Submissions, while granting Suno a worldwide royalty-free license (including creating derivative works); the terms state this license covers using content 'to train, develop, fine-tune or otherwise improve' its models, with no training opt-out documented. Training other models on outputs is prohibited.
适合谁Who it suits
做 AI 音乐、在意「上传素材归我」的人。商用要走付费下载通道（见许可列），Remix 一律非商用。For AI-music makers who care that uploads stay theirs. Commercial exploitation runs through paid download channels (see the license column); Remixes are non-commercial across the board.
商用许可Commercial use
官方 ToS：Pro / Premier 付费档——「Suno hereby assigns to you all of its right, title and interest in and to any Output owned by Suno」；免费 / Basic 档——输出仅限「lawful, personal and non-commercial purposes」。商用许可绑定「Permitted Commercial Use」：只有经付费档下载配额下载的 Output 可商用；Remix（他人二创）无论档位一律非商用。Official ToS: on the Pro / Premier paid tiers, 'Suno hereby assigns to you all of its right, title and interest in and to any Output owned by Suno'; on the free / Basic tier, Outputs may only be used for 'lawful, personal and non-commercial purposes'. Commercial rights are tied to 'Permitted Commercial Use' — only Outputs downloaded through your paid tier's download allocation may be commercially exploited. Remixes are non-commercial regardless of tier.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：1 条官方链接 · 核验 2026-10-20 · verifiedSources: 1 official links · checked 2026-10-20 · verified

ElevenLabsElevenLabs

ElevenLabs · 轻档案 · 不上八维light entry · not on the 8-axis grid

国外出品foreign-made 大陆可直连direct OK 未见于备案名单not in filing list

能做什么What it does
语音合成与配音（TTS / 朗读 / Voice Design 声音设计 / 克隆）；网页、API 与移动端。Text-to-speech, dubbing and voice design / cloning across web, API and mobile.
数据去哪Where your data goes
官方 ToU §4(d) 授予平台为运营与改进（含训练）使用内容的许可；§4(i) 提供 Data use 设置菜单可退出把内容用于训练。声音克隆须本人授权或取得合法授权。Per the official ToU §4(d), users grant the platform a license to use content for operating and improving the service (including training); §4(i) provides a Data use settings menu to opt out of training use. Voice cloning requires the person's own consent or lawful authorization.
适合谁Who it suits
做配音、有声内容、语音产品的团队。训练退出是自助菜单（相对少见，值得点名）；克隆他人声音前先拿授权。For teams doing dubbing, audio content or voice products. The training opt-out is a self-serve menu (uncommon — worth knowing); get consent before cloning anyone's voice.
商用许可Commercial use
官方 ToU §1(c)：付费用户可将输出用于商用，免费用户仅限「non-commercial purposes」；§4(c)(ii)：用户保留 Output 的权利与所有权。商用权与档位绑定——换档即换商用许可。Official ToU §1(c): paid users may use Outputs commercially, while free users are limited to 'non-commercial purposes'. §4(c)(ii): users retain rights and ownership in their Outputs. Commercial rights are tier-bound — changing plans changes the commercial license.

多少钱：价格常变，本站不存报价，请看 官方定价页。Price: prices change often; this site stores no quotes — see the official pricing page.

出处：1 条官方链接 · 核验 2026-10-20 · verifiedSources: 1 official links · checked 2026-10-20 · verified

深水指南Deep guides

产品档案回答「这个工具是什么」；深水指南回答「这类工程问题怎么办」。六篇各管一件事：编排、记忆、规划、评测、成本、护栏——讲机制与取舍，不做框架排名。每篇的关键事实都挂官方文档或论文出处。Product entries answer "what is this tool"; the deep guides answer "how do you handle this class of engineering problem". Six guides, one topic each — orchestration, memory, planning, evaluation, cost, guardrails — covering mechanisms and trade-offs, with no framework rankings. Every key fact carries an official-doc or paper source.

记忆系统Agent memory

agent 记忆系统Agent memory systems

会话内的状态、跨会话的持久事实、外部记忆库——三层记忆各自解决什么问题，以及共同的坑。In-session state, cross-session facts, and external stores — what each memory layer solves, and the pitfalls they share.

 成本控制Cost control

agent 的成本控制Cost control for agents

钱花在哪、怎么省：token 记账、前缀缓存、上下文膨胀与步数上限——只讲机制，不报价格。Where the money goes and how to curb it: token accounting, prefix caching, context bloat, and step caps — mechanisms only, no prices.

 评测Evaluation

agent 评测Evaluating agents

没有评测就没有迭代：上线前用数据集测回归，上线后盯住真实流量，以及评测器自己会犯的坑。No evaluation, no iteration — offline regression testing, live monitoring, and the pitfalls of evaluators themselves.

 护栏与安全Guardrails & safety

agent 的护栏Guardrails for agents

输入输出与工具三层防线、触发即止的 tripwire 机制，以及并行与阻塞的成本取舍。Input, output, and tool-level defenses, tripwires that halt on trigger, and the cost trade-off between parallel and blocking checks.

 多 agent 编排Multi-agent orchestration

多 agent 编排Multi-agent orchestration

多个 agent 怎么分工、怎么交接、怎么收拢结果，以及什么时候根本不该上多 agent。How multiple agents divide work, hand off, and merge results — and when not to go multi-agent at all.

 规划Planning

agent 的规划Planning for agents

先想清楚步骤再动手，还是边做边定——规划模式怎么选，以及规划为什么会失效、失效了怎么救。Plan first or decide as you go — how to choose a planning pattern, and why plans fail.

先回答一件事：你到底缺什么能力Answer one question first: what capability are you actually missing

MCP 不是「让 AI 更聪明」，是给 AI 接上它本来够不到的东西。如果你要的只是聊天和写代码，你不需要它。你装，它调——10 个条目：7 个是 MCP 官方 reference server，2 个由各自的官方团队维护（Playwright / Context7），1 个是厂商官方（GitHub 自己的 github-mcp-server）。MCP is not "making AI smarter" — it is giving AI access to what it could not otherwise reach. If all you want is chat and writing code, you do not need it. You install it, the AI calls it — 10 entries: 7 are the MCP org's reference servers, 2 are maintained by their own official teams (Playwright / Context7), and 1 is vendor-official (GitHub's own github-mcp-server).

  1. 1我要读写本地文件 / 操作 GitI need to read/write local files / work with Git
    filesystem、gitfilesystem, git两个都是读写权限，git 那一类值得先看 scope 里能限制到哪些目录。Both have read and write permissions; for the git category it is worth checking what directories the scope can be limited to.
  2. 2我要它记住东西、跨会话召回I need it to remember things across sessions
    memorymemory这是最容易被误解的一类——它不是「模型记忆」，是外部知识图谱存储。This is the most misunderstood category — it is not "model memory" but an external knowledge-graph store.
  3. 3我要让 AI 上网拿资料I need AI to reach the web for material
    context7、fetchcontext7, fetch风险最高的一类是 fetch：档案里引了官方警告——可访问本地与内网 IP，无地址白名单、无审批机制。fetch carries the highest risk: the entry quotes an official warning — it can reach local and internal IPs, with no address allowlist and no approval mechanism.
  4. 4我要控制真实浏览器I need to drive a real browser
    playwrightplaywright注意官方明确声明来源限制不是安全边界，需要真隔离得用 --isolated 或容器化网络策略。Note the official explicit statement that source restrictions are not a security boundary; real isolation requires `--isolated` or a containerised network policy.
  5. 5我要点时间 / 拆解推理 / 跨平台搜文件I want time / decomposition / cross-platform file search
    time、sequential-thinking、everythingtime, sequential-thinking, everything三个都是只读，风险最低。All three are read-only, so risk is lowest.
  6. 6我要操作 GitHub（仓库 / Issue / PR / Actions）I need to work with GitHub (repos / issues / PRs / Actions)
    看 github-mcpSee github-mcpGitHub 官方自营的 MCP server，与归档的旧参考 server 不是同一条线；权限由 GitHub token 决定，可收紧到只读。GitHub's own MCP server — not the same line as the archived reference server; permissions follow the GitHub token and can be tightened to read-only.

⚠ 一个容易忽略的维度：数据出境。Context7 默认以远程服务交付，查询内容会发到 Upstash 服务器；GitHub MCP Server 提供远程托管与本地 Docker 两种形态；其余八个都是本地 stdio 进程。另外——装 MCP 需要你手动做，调用不需要：server 启动后自动把工具清单报给你的客户端，之后调不调、什么时候调，由 agent 自己判断，你只在审批环节介入。⚠ An easily overlooked dimension: data leaving the country. Context7 is delivered as a remote service by default — queries go to Upstash's servers; GitHub MCP Server offers both a remote hosted and a local Docker form; the other eight are local stdio processes. And — installing an MCP server is manual; calling it is not: once running, the server reports its tool list to your client automatically, and after that the agent decides whether and when to call it. You only get involved at the approval step.

全部对象All objects

按统一坐标系排列。点开任一对象可看到 模型与开放条件 / 运行位置 / 本地文件 / 关机后的任务 / 工具与扩展 / 上下文与记忆 / 权限与限制 / 适合什么任务 / 传输方式 / 认证机制 / 权限范围 的完整记录、证据链接与未知项清单。Everything on one coordinate system. Open any object to see its full record across Model & openness / Where it runs / Local files / Tasks after shutdown / Tools & extensions / Context & memory / Permissions & limits / Best for / Transport / Authentication / Permission scope, with source links and a list of unknowns.

Context7Context7

Upstash

工具分区里唯一的远程托管条目 —— 为 Agent 提供版本感知的库文档，MIT 开源但核心服务在 Upstash 侧。The only remotely hosted entry in the MCP track — version-awar

// line 已是双节点，勿走 esc -->
部分核验Partly verified 详情 →Details →

Everything MCP ServerEverything MCP Server

Model Context Protocol

不是可用工具——官方明说这是给 MCP 客户端开发者做功能测试的 server，唯一用途是验证客户端协议实现是否正确。Not a usable tool — the official page states plainly that this

// line 已是双节点，勿走 esc -->
已核验Verified 详情 →Details →

Fetch MCP ServerFetch MCP Server

Model Context Protocol

把网页内容转成 Markdown 的 reference server，官方在其 README 里明确标注了安全警告：它能访A reference server that converts web pages to Markdown, with a

// line 已是双节点，勿走 esc -->
已核验Verified 详情 →Details →

Filesystem MCP ServerFilesystem MCP Server

Model Context Protocol

MCP 官方 reference server 里的文件操作实现。它是「白名单目录」式访问：必须显式声明允许目录，不允许静默The file-operation implementation among MCP's official referen

// line 已是双节点，勿走 esc -->
已核验Verified 详情 →Details →

Git MCP ServerGit MCP Server

Model Context Protocol

MCP 官方 reference server 里的版本控制实现。它作用域限定在指定 git 仓库，但具备完整的本地写操作能The version-control implementation among MCP's official refere

// line 已是双节点，勿走 esc -->
已核验Verified 详情 →Details →

GitHub MCP ServerGitHub MCP Server

GitHub

GitHub 官方自己的 MCP server（仓库自述 "GitHub's official MCP Server"）， GitHub's own official MCP server — not the archived MCP-org re

// line 已是双节点，勿走 esc -->
部分核验Partly verified 详情 →Details →

Playwright MCP ServerPlaywright MCP Server

Microsoft

微软维护的浏览器自动化 server。它用无障碍树而非截图驱动，不需要视觉模型。输出是结构化快照。A browser automation server maintained by Microsoft, driven by

// line 已是双节点，勿走 esc -->
已核验Verified 详情 →Details →

Time MCP ServerTime MCP Server

Model Context Protocol

只做两件事：查当前时间、换时区。无文件无网络，是权限风险最低的一类 server。A reference server providing date and time information to agen

// line 已是双节点，勿走 esc -->
已核验Verified 详情 →Details →

本站目前是「核验快照」，还没有实测数据This site is currently a verification snapshot — no measured runs yet

下面每个 server 的权限范围、传输方式与工具清单都能回到官方 README 核对，但我们还没有在真实客户端里跑过它们。Every server's permission scope, transport and tool list below can be traced back to its official README, but we have not yet run them inside a real client. 实测协议已经写好（统一任务、统一验收、记录人工介入与返工次数），任务清单也已就绪，The measurement protocol is written (shared tasks, shared acceptance criteria, recording human interventions and rework counts) and the task list is ready, but it is 尚未执行。not yet run.

MCP server 的风险不在「能不能干活」，而在权限边界是否清楚、越界是否被拒。The risk in an MCP server is not whether it works, but whether the permission boundary is clear and overreach is refused. 真正的选型请用你自己的输入、预算与验收标准跑一遍。For an actual choice, run them yourself against your own inputs, budget and acceptance criteria.

11 个维度，每格都能回溯11 dimensions, every cell traceable

所有对象在同一坐标系下被描述（MCP 服务器另有 3 个特有维度）。All objects are described on the same coordinate system (MCP servers have 3 additional dimensions). 点开任一对象，每一格下面都有官方源链接与核验日期——不认同可以自己回去查。Open any object and every cell carries an official source link and a checked date — disagree and verify it yourself.

01

模型与开放条件Model & openness

用谁的模型？能不能换第三方？多模型路由怎么配？Whose model does it use? Can you switch to a third party? How is multi-model routing configured?

02

运行位置Where it runs

本地 / 云端 / 混合？索引在本地还是云端？远程控制不等于云端执行。Local, cloud or hybrid? Is the index local or remote? Remote control is not the same as cloud execution.

03

本地文件Local files

能访问多大范围？未打开的大文件怎么读？符号级能力如何？How wide an area can it reach? How does it read large files you have not opened? How capable is it at the symbol level?

04

关机后的任务Tasks after shutdown

电脑关了还能继续跑吗？这是本地形态最常见的误解来源。Does it keep running once the computer is off? This is the most common misconception about local forms.

05

工具与扩展Tools & extensions

工具链有多宽：MCP · LSP · 插件 · Hooks · 自定义指令。How wide is the toolchain: MCP · LSP · plugins · hooks · custom instructions?

06

上下文与记忆Context & memory

索引算法、上下文窗口、会话恢复与分叉、跨会话记忆边界。Indexing algorithm, context window, session resume and fork, and the boundary of cross-session memory.

07

权限与限制Permissions & limits

沙箱与审批边界、凭据管理、数据是否出境、额度如何计算。Sandbox and approval boundaries, credential handling, whether data leaves your machine, and how quota is charged.

08

适合什么任务Best for

最擅长什么任务？什么情况下别用它？What tasks is it best at? When should you not use it?

09

传输方式Transport

客户端怎么连上它：stdio / SSE / Streamable HTTP / 远程。How the client connects to it: stdio / SSE / Streamable HTTP / remote.

10

认证机制Authentication

怎么证明身份。注意认证 ≠ 授权，前者是「你是谁」，后者是「你能做什么」。How identity is proved. Note that authentication is not authorisation: the former is who you are, the latter is what you may do.

11

权限范围Permission scope

能碰到多少东西？能改吗？能删吗？能收紧吗？这是 MCP 分区第一优先级。How much can it reach? Can it modify? Delete? Be tightened? This is the first priority in the MCP track.

全部数据一览All data at a glance

同一张表横向对比。点对象名进详情页看逐格证据。The same table side by side. Click an object name to see per-cell evidence.

对象Object形态Form厂商Vendor适合Best for核验日Checked状态Status
C7Context7Context7 MCPMCP Upstash 需要查当前版本库文档的场景：库升级后的 API 变更、 新版本引入的配置项、模型不确定的第三方库用法。 不适合查内部代码或私有库。Scenarios needing current-version library documentation: API changes a 2026-10-01 partial
EEverything MCP ServerEverything MCP Server MCPMCP Model Context Protocol 仅适合一件事：测试你的 MCP 客户端是否正确实现了协议。 不适合任何实际任务场景。Scenarios where you are testing an MCP client implementation rather th 2026-10-08 verified
FFetch MCP ServerFetch MCP Server MCPMCP Model Context Protocol 把公开网页内容转成 Markdown 供 Agent 读取。 不适合在能访问内网的环境中使用， 也不适合需要登录态的页面。Scenarios needing to pull a web page's content as Markdown for an agen 2026-10-08 verified
FSFilesystem MCP ServerFilesystem MCP Server MCPMCP Model Context Protocol 让 Agent 安全读写指定目录内的文本与媒体文件。 不适合需要删除操作或需要跨目录聚合的场景。Scenarios where an agent must read or write within explicitly declared 2026-10-08 verified
GGit MCP ServerGit MCP Server MCPMCP Model Context Protocol 让 Agent 读取仓库状态、历史与 diff，并在受控前提下执行提交与分支操作。 不适合需要 push / PR / 远程仓库管理的场景。Scenarios where an agent should make version-control operations inside 2026-10-08 verified
GHGitHub MCP ServerGitHub MCP Server MCPMCP GitHub 一句话：让任意 MCP 客户端用自然语言操作 GitHub。 客户端包括 VS Code、Claude Desktop、Cursor、WinAnyone who wants their MCP client (VS Code, Claude Desktop, Cursor, Wi 2026-10-08 partial
MKnowledge Graph Memory ServerKnowledge Graph Memory Server MCPMCP Model Context Protocol 让 Agent 跨会话记住用户偏好、项目背景等长期信息。 不适合存放需要审计或追溯的场景（图谱会原地修改）。Scenarios needing cross-session memory that survives restarts and is q 2026-10-08 verified
PPlaywright MCP ServerPlaywright MCP Server MCPMCP Microsoft 需要浏览器自动化且要保持会话状态的场景： 探索性自动化、自愈测试、长时自主工作流。 不适合高吞吐编码 agent——README 自己建议这Scenarios where an agent must operate a real browser and read the page 2026-10-08 verified
SSequential Thinking MCP ServerSequential Thinking MCP Server MCPMCP Model Context Protocol 让 Agent 用结构化、可修订、可分支的方式推进复杂问题的分析。 不适合简单任务（增加调用轮次无收益）。Scenarios where a complex problem benefits from being worked through s 2026-10-08 verified
TTime MCP ServerTime MCP Server MCPMCP Model Context Protocol 给 Agent 提供可靠的时间与时区换算能力。 适合需要避免模型猜时间的场景（如日程、跨时区协作）。Scenarios where an agent needs the current date and time to reason abo 2026-10-08 verified

为什么这里不给总分排名Why there is no overall ranking

这不是省略，是方法上的明确取舍。This is not an omission but a deliberate methodological choice.

  • 缺统一实测，就不给分。跨工具的跑分口径不同——题目、预算、执行环境都不一样，A 的 90 分和 B 的 88 分不可比。No shared measurement, no score. Benchmarks across tools use different rules — different questions, budgets and environments. A 90 and a B 88 are not comparable.
  • 「未知」是合法答案。"Unknown" is a legitimate answer.查不到就写「未知」并说明为什么，不用推测填充。缺信息本身也是信息。
  • 来源链接、类型与核验日都在详情页里，可以逐条回去复核。The source link, its kind and the checked date are all on the detail page, so you can walk any of them back and re-check.
  • lifecycle 单独标注。核验日新不等于数据有效——上游归档后数据会失效，所以单独标 active / maintenance / archived。lifecycle is marked separately. A recent check date does not mean the data is still valid — it goes stale when upstream archives the repo, so active / maintenance / archived are tracked separately.
  • 不同层不混排。成品 agent、自己搭的底座、给 agent 装的工具是三种角色，坐标系不同；硬合成一张 33 行的表只会制造假的可比性。Layers are not mixed. Finished agents, bases you assemble yourself, and tools you install for an agent are three different roles with different coordinate systems; forcing them into one 33-row table would manufacture false comparability.

可信度标记怎么读How to read the confidence marks

详情页与上表都带Both the detail pages and the table above carry confidence，它反映, which reflects 本站数据的完整度how complete this site's data is，不是对产品的评价。, not a judgement of the product.

verified11 个维度均有官方源支撑，且核验日在 90 天内all 11 dimensions have official sources and were checked within 90 days
partial部分维度标为未知，或官方文档不可访问，或核验日超过 90 天some dimensions are marked unknown, or the official docs were unreachable, or checked over 90 days ago
stale官方已发布重大变化，本站尚未核验the vendor has shipped a significant change that this site has not re-checked

本站当前的完整度Current completeness on this site

数据层用 node scripts/audit-gaps.mjs 可随时复核这个数字。 以下是最近一次核验的快照（核验日 2026-09-29）：

形态Form已补齐Filled说明Meaning
tools 110/110 官方 README 信息充分，绝大多数维度可直接引述Official READMEs cover most of it, so most dimensions can be quoted directly

全站合计 415/446 维度已补齐Site-wide 415/446 dimensions filled（93%）

剩余未补齐项分三类：官方未公开（索引算法、沙箱实现细节本就不对外说明）、需实测才能确定（大仓库表现、CI 无 TTY 行为）、客观渠道不可达。The remaining gaps fall into three kinds: not disclosed by the vendor (indexing algorithms, sandbox internals are simply not documented), requires measurement (behaviour on very large repositories, CI without a TTY), and unreachable through objective channels.

我们选择留白而不是填「已支持」——错误的成本最终由使用者承担。We chose to leave a blank rather than fill in "supported" — the cost of an error is ultimately borne by whoever uses it.

数据来源与复核方式Sources and how to re-check them

全部数据来自开源仓库All data comes from the open-source repository speculcom/ai-agent-guide（CC BY 4.0）。

每个条目都是纯 Markdown + frontmatter，含 8 个维度、证据链Every entry is plain Markdown plus frontmatter, carrying 8 dimensions and an evidence chain接、核验日与可信度标记。仓库内含：

  • METHODOLOGY.md —— 方法论总纲与四条核心规则
  • axes/ —— 每个维度的定义与判定标准（含正反例）
  • SCHEMA.md —— 数据格式规范与构建校验规则
  • tracks/*/tasks/_protocol.md —— 实测协议(尚未执行not yet run，故本站暂无实测结论, so this site has no measured conclusions)

发现错误或有新证据，欢迎提 Issue 或 PR——每条修正都会注明依据与影响范围。Found an error or have new evidence? Open an issue or a PR — each correction notes its basis and blast radius.