← MCP

Context7Context7

Upstash

检索Retrieval 网络Network 远程Remote partial lifecycle: active 核验Checked 2026-10-23

远程托管的文档服务 —— 为 Agent 提供版本感知的库文档，MIT 开源但核心服务在 Upstash 侧。

A remotely hosted documentation service — version-aware library documentation for agents; MIT-licensed, but the core service lives on Upstash's side.

适合与不适合When to use it, when not to

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。

不适合涉密场景——查询内容会发送到 Upstash 服务器，需先评估数据出境。也不适合查内部代码或私有库。

Scenarios needing current-version library documentation: API changes after an upgrade, configuration options introduced in a new version, usage of third-party libraries the model is unsure about.

Not a fit for confidential material — queries are sent to Upstash servers, so assess data leaving your machine first.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网仓库在 GitHub，大陆直连波动（同日实测 200/000 混合）；npm / PyPI 安装源实测稳定可达。Hosted on GitHub — direct reachability from mainland fluctuates (200/000 mixed on the same day); npm / PyPI install routes measured stable.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

11 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 本身不调用模型，是被客户端调用的接口实现。为 Agent 提供版本感知的库与 API 文档，解决模型训练数据滞后导致的使用错误。
It calls no model itself; it is an interface implementation called by clients. It gives agents version-aware library and API documentation, solving usage errors caused by stale model training data.
运行位置Where it runs: 远程托管 HTTP 服务（https://mcp.context7.com/mcp），

由 Upstash 运营。这是本分区唯一默认走远程的条目，其余 reference server 均为本地 stdio 进程。
A remotely hosted HTTP service (https://mcp.context7.com/mcp), operated by Upstash. This is the only entry in this section that defaults to remote; all other reference servers are local stdio processes.
本地文件Local files: 无任何本地文件访问能力。纯文档检索服务。
No local file access at all. A pure documentation-retrieval service.
关机后的任务Tasks after shutdown: 作为远程托管服务始终可用，与客户端进程生命周期无关—— 这是本分区唯一不受本机开关影响的条目。

但服务可用性取决于 Upstash 侧的服务状态，本仓库的 releases 不能保证服务端可用性，两者需分开判断。
Always available as a remotely hosted service, independent of the client process lifecycle — the only entry in this section unaffected by your machine's switch.

But service availability depends on Upstash's service status; this repository's releases cannot guarantee the server side — the two must be judged separately.
工具与扩展Tools & extensions: 提供 MCP 工具用于检索库文档。

README 强调正确用法：使用 Library Id、指定版本（Specify a Version）、以及通过 rule 声明「需要库文档时总是使用 Context7」。输出为结构化文档内容。
It provides MCP tools for retrieving library documentation.

The README stresses correct usage: use Library Ids, specify a version ("Specify a Version"), and declare via a rule that Context7 should always be used when library docs are needed. Output is structured documentation content.
上下文与记忆Context & memory: 无跨会话记忆，但文档检索本身是"外部记忆"—— 弥补模型训练数据滞后，这是它的核心定位。
No cross-session memory, but documentation retrieval is itself "external memory" — compensating for stale model training data; that is its core positioning.
权限与限制Permissions & limits: 权限风险与本地 server 完全不同：不碰本地文件，但要把查询内容发送到 Upstash 服务器。

传入 Library Id 与具体问题即产生网络请求。认证用 API Key（Authorization: Bearer YOUR_API_KEY）。无独立审批机制。
The permission risk profile is completely different from local servers: it touches no local files, but your query content is sent to Upstash servers.

Passing a Library Id and a question produces a network request. Authentication uses an API key (Authorization: Bearer YOUR_API_KEY). There is no separate approval mechanism.
适合什么任务Best for: 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 where you need current-version library docs: API changes after a library upgrade, config options introduced by new versions, third-party library usage the model is unsure about. Not a fit for internal code or private libraries.
传输方式Transport: 远程 Streamable HTTP，服务端点为 https://mcp.context7.com/mcp。

本分区唯一以远程服务为默认形态的条目—— 其余官方 reference server 均为本地 stdio 进程。
Remote Streamable HTTP, server endpoint https://mcp.context7.com/mcp.

The only entry in this section whose default form is a remote service — all other official reference servers are local stdio processes.
认证机制Authentication: 需 API Key，通过 Authorization: Bearer YOUR_API_KEY 头传递。 README 说明可经 OAuth 流程由安装器自动生成 key。

额度档位已核验（2026-10-23 滚动复验，官方 plans 页）： Free 1,000 次/月（Search API + Context API 合计），

Pro 2,000 次/席位/月、超出不阻塞（超量单价按本站的报价记录约定不存，见官方 plans 页）。

另Free 档本身就用 OAuth 2.0（官方对比表把 OAuth 2.0 列为 Free 的能力）。 ⚠ Free 超出额度后的确切行为官方页面未给答案正文，只列了问题标题。
An API key is required, passed via the Authorization: Bearer YOUR_API_KEY header. The README says installers can generate the key automatically through an OAuth flow.

Quota tiers verified (rolling re-check 2026-10-23, official plans page): Free 1,000 calls/month (Search API + Context API combined),

Pro 2,000 calls/seat/month, overage does not block (overage unit price not stored per our price-recording policy; see the official plans page).

Also, the Free tier itself already uses OAuth 2.0 (the official comparison table lists OAuth 2.0 under Free). ⚠ The exact behaviour after Free exceeds its quota: the official page gives no answer body, only the question title.
权限范围Permission scope: 无本地权限风险，但有数据出境风险：查询内容会发送到 Upstash 服务器。传入的库名与问题文本属于对外传输内容，涉密场景需评估。

已核到一条官方对隐私的明确承诺（官方 FAQ 标题即 "How do you keep the privacy of my queries?"）：

查询只发文档问题与库名/ID，不发完整 prompt、源代码与对话历史；且官方要求agent 不要在查询里带凭据、个人数据或专有代码。

服务端自身的文档索引范围由 Upstash 掌握，客户端不可见。
No local permission risk, but there is a data-egress risk: query content is sent to Upstash servers. The library names and question text you pass are outbound content; sensitive scenarios need an assessment.

The 2026-10-08 check adds one explicit official privacy commitment (the official FAQ title is "How do you keep the privacy of my queries?"):

Queries send only the documentation question and the library name/ID, not full prompts, source code or conversation history; and the official docs require agents not to include credentials, personal data or proprietary code in queries.

The document-index scope on the server side is Upstash's; it is not visible to clients.

头号误解Top misconception

以为是纯本地工具，实际查询会发到 Upstash 的远程服务器Assuming it is a purely local tool — queries actually go to Upstash's remote servers
不指定版本号会拿到不匹配的文档，README 专门强调Specify a VersionOmitting a version number gets you mismatched docs — the README calls out "Specify a Version" specifically
以为开源就等于全部本地，代码 MIT 但核心是托管服务Assuming open source means fully local — the code is MIT but the core is a hosted service

未知项清单Known unknowns

Free 档超出 1,000 次/月后的确切行为（官方页面只列了问题标题，未给答案正文；不做推测）
Enterprise 档的具体内容（报价按本站的报价记录约定本站不采）
文档索引的覆盖范围与更新频率（官方有「Keeping Libraries Fresh」专题，未取正文）
服务端 SLA
官方隐私声明（只发查询与库名、不发源码与对话历史）未做实测验证

The exact behaviour after the Free tier exceeds 1,000 calls/month (the official page lists only the question title, no answer body; no speculation)
What the Enterprise tier includes (prices not collected per our price-recording policy)
Document-index coverage and refresh frequency (the official docs have a "Keeping Libraries Fresh" topic; its body was not captured the 2026-10-08 check)
Server-side SLA
The official privacy statement (queries and library names only; no source code or conversation history) is not verified by measurement

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-01。

类型Type	名称Name	链接Link
docs	Upstash · Context7 README（API Key 与远程端点说明、「本仓只托管 MCP server 源码，API 后端/解析/爬虫为私有」的自述）Upstash · Context7 README (API-key and remote-endpoint notes, and the self-description that "this repo hosts only the MCP server source; the API backend/parsing/crawlers are private")	https://github.com/upstash/context7
pricing	官方 Pricing & Plans（滚动核验 2026-10-23：Free 1,000 次每月、Pro 2,000 次每席位每月且超出不阻塞、Private Repo Parsing 是独立计费项、Enterprise Custom 含 SOC-2/SSO/Self-Hosted；报价按本站的报价记录约定不存；与第三方「Pro 5,000 次」冲突，按官方 2,000）Official Pricing & Plans (rolling verification 2026-10-23: Free 1,000 calls/month, Pro 2,000 calls per seat per month with no blocking on overage, Private Repo Parsing as a separate billing item, Enterprise Custom including SOC-2/SSO/Self-Hosted; prices not stored per our price-recording policy; conflicts with a third-party "Pro 5,000 calls" claim — official 2,000 adopted)	https://context7.com/docs/plans-pricing.md
changelog	Upstash · Context7 Releases（[email protected]）Upstash · Context7 Releases ([email protected])	https://github.com/upstash/context7/releases
changelog	Upstash · Context7 CommitsUpstash · Context7 Commits	https://github.com/upstash/context7/commits/main
docs	Model Context Protocol · 规范Model Context Protocol · specification	https://modelcontextprotocol.io/specification

实测记录Measured runs

本档未在真实客户端逐项跑通；工具清单与权限来自官方 README。

实测建议：这个 server 的实测重点与其他条目不同—— 要测的是文档准确度（给定库与版本，返回的文档是否匹配当前 API），而非能否调用成功。

This site has not completed measured runs. The test protocol is at protocols/mcp/_protocol.md.

Suggested measurement: the focus for this server differs from the other entries — what to test is documentation accuracy (given a library and version, do the returned docs match the current API), not whether the call succeeds.

为什么它和其他条目不一样Why it is unlike the other entries

维度	其余 reference server	Context7
运行形态	本地进程（stdio）	远程 HTTP 服务
数据流向	本机内	查询发到 Upstash
认证	无（靠进程权限）	需 API Key
额度	无	有（官方提"更高额度"）
开源许可	参考实现，无商业形态	MIT 代码 + 托管服务

这带来的是完全不同的一类风险：其他 server 的风险是「能碰你的什么」， Context7 的风险是「你的查询会发到哪里去」。

Dimension	Other reference servers	Context7
Run form	Local process (stdio)	Remote HTTP service
Data flow	Stays on your machine	Queries go to Upstash
Authentication	None (relies on process permissions)	API key required
Quota	None	Yes (the official docs mention "higher quotas")
Open-source licence	Reference implementation, no commercial form	MIT code + hosted service

This brings a completely different class of risk: for other servers the risk is "what can it touch of yours"; for Context7 the risk is "where do your queries go".

变更记录说明Change-log notes

本对象有独立 releases：

版本	日期
`[email protected]`	2026-09-22
`@upstash/[email protected]`	2026-09-22
`@upstash/[email protected]`	2026-09-22

一个仓库发多个包，版本号不同步，引用时要注意包名。

This object has its own releases:

Version	Date
`[email protected]`	2026-09-22
`@upstash/[email protected]`	2026-09-22
`@upstash/[email protected]`	2026-09-22

One repository publishes multiple packages, with unsynchronised version numbers — watch the package name when referencing.

正确用法（README 强调的三点）Correct usage (the three points the README stresses)

这是最有操作价值的信息：

要点	说明
Use Library Id	用库 ID 而非模糊名字查询
Specify a Version	指定版本 —— 不指定会拿到不匹配的文档
Add a Rule	声明规则：需要库文档时总是使用 Context7

「Specify a Version」这一条尤其重要—— 模型训练数据滞后是这类工具要解决的核心问题，如果不指定版本，很可能拿到的是新版本文档，行为却按老版本写。

This is the most operationally valuable information:

Point	Explanation
Use Library Id	Query by library ID, not fuzzy names
Specify a Version	Specify the version — without it you get mismatched docs
Add a Rule	Declare a rule: always use Context7 when library docs are needed

"Specify a Version" matters most — stale model training data is the core problem this class of tool solves; if you do not pin the version, you likely get docs for a newer version while the code behaves per the old one.

认证方式与额度：三档，两处容易看错Authentication and quotas: three tiers, two easy misreadings

手动配置：使用服务端点 https://mcp.context7.com/mcp
并通过 Authorization: Bearer YOUR_API_KEY 头传递 key

安装器支持 OAuth 流程自动生成 key。Free 档本身即支持 OAuth 2.0。

官方原话：API Key Recommended —— 获取免费 key 以获得更高额度。

额度档位（2026-10-01 核验官方 plans 页）

（报价按本站的报价记录约定不存，一律看官方 plans 页；额度是能力边界，照录。）

档	含额度	超出后
Free	1,000 次/月	阻塞（具体行为官方未明示）
Pro	2,000 次 / 席位	不阻塞（超量按用量另计，单价见官方页）
Enterprise	Custom	Custom

两处最容易看错的地方：

① Pro 的「Unlimited API Calls」不是「不限量免费」，是「不阻塞」。 官方对比表把 Pro 的 Additional API Calls 一栏写成 "Unlimited（超量按用量计费）" —— 前半句是不阻塞，后半句意味着超量要另付钱。

② Private Repo Parsing 是完全独立的计费项。 它与 API 调用额度是两个钱包，Pro 的席位价并不包含它。要把私有库文档接进来，这笔钱要单独算（单价见官方页）。

第三方数字与官方不一致（已按官方为准）

多个第三方站（2026-09 前后）称 Pro 是「5,000 次/席位」—— 官方页面是 2,000。本档案采官方，第三方数字视为旧制或臆测。

开源的边界要划清

官方 README 自己说明：这个仓只托管 MCP server 的源码， API 后端、解析引擎、爬取引擎都是私有的。 所以「MIT」指的是那部分客户端代码，不是整个服务。

Manual config: use the server endpoint https://mcp.context7.com/mcp
and pass the key via the Authorization: Bearer YOUR_API_KEY header

Installers support an OAuth flow that generates the key automatically. The Free tier itself supports OAuth 2.0.

Official wording: API Key Recommended — get a free key for higher quotas.

Quota tiers (verified against the official plans page, 2026-10-01)

(Price figures are not stored per our price-recording policy — always see the official plans page; quotas are capability boundaries and are recorded as-is.)

Tier	Included quota	After exceeding
Free	1,000 calls/month	Blocking (exact behaviour not stated officially)
Pro	2,000 calls / seat	Does not block (overage billed by usage; unit price on the official page)
Enterprise	Custom	Custom

The two easiest misreadings:

① Pro's "Unlimited API Calls" does not mean "unlimited and free" — it means "does not block". The official comparison table writes Pro's Additional API Calls column as "Unlimited (overage billed by usage)" — the first half means no blocking, the second half means overage costs extra.

② Private Repo Parsing is a completely separate billing item. It is a different wallet from the API-call quota; the Pro seat price does not include it. If you want private-library docs ingested, that cost is calculated separately (unit price on the official page).

Third-party figures disagree with the official page (official prevails)

Several third-party sites (around 2026-09) claim Pro is "5,000 calls/seat" — the official page says 2,000. This entry follows the official figure; the third-party number is treated as outdated or speculative.

Draw the open-source boundary clearly

The official README itself states: this repository hosts only the MCP server's source code; the API backend, parsing engine and crawling engine are proprietary. So "MIT" refers to that client-side code, not the whole service.

风险评估Risk assessment

风险类型	评价
本地文件泄露	无 — 不碰本地文件
内网探测	无 — 不访问任意 URL
数据出境	有，但范围窄 — 只发文档问题 + 库名/ID；不发完整 prompt、源代码、对话历史（官方声明，见下方说明）
凭据风险	低 — 只用一个 API Key，无文件权限；官方还要求 agent 不要在查询里带凭据
服务可用性	依赖 Upstash 服务，本仓库无 releases 可保证服务状态
预算失控	有 — Free 超 1,000 次/月后行为不明；Pro 超出不阻塞、按用量另计；私有库解析是独立计费项（单价见官方页）

涉密场景要评估的部分：如果你查的库名、报错信息、代码片段本身敏感，这些会离开本机。

但已核到一条官方对隐私的明确承诺（官方 FAQ 设了专题 "How do you keep the privacy of my queries?"）：

查询只发送文档问题与库名/ID， 不发送完整 prompt、源代码与对话历史；官方并要求 agent 不要在查询里携带凭据、个人数据或专有代码。

⚠ 这条声明是官方自述，本站未做端到端实测验证 —— 即没有抓包确认发出去的内容真如所述。

Risk type	Assessment
Local file leakage	None — touches no local files
Internal-network probing	None — does not visit arbitrary URLs
Data egress	Yes, but narrow in scope — sends only the documentation question + library name/ID; not full prompts, source code or conversation history (official statement, see note below)
Credential risk	Low — one API key only, no file permissions; the official docs also require agents not to put credentials in queries
Service availability	Depends on the Upstash service; this repository's releases cannot guarantee service status
Budget overrun	Yes — behaviour after Free exceeds 1,000 calls/month is unknown; Pro does not block on overage but bills by usage; private-repo parsing is a separate billing item (unit price on the official page)

The part sensitive scenarios must assess: if the library names, error messages or code snippets you query are themselves sensitive, they leave your machine.

But the 2026-10-08 check adds one explicit official privacy commitment (the official FAQ has a dedicated entry, "How do you keep the privacy of my queries?"):

Queries send only the documentation question and the library name/ID, not full prompts, source code or conversation history; the official docs also require agents not to carry credentials, personal data or proprietary code in queries.

⚠ This statement is the vendor's own account; this site has not verified it by measurement — i.e. no packet capture confirmed the outbound content matches the description.