Token导航 LogoToken导航TokenDH.com
研究检索敏感数据github未标认证来源可访问许可证需确认审计提醒

mcp-oauth-setupMCP OAuth 设置

Agent Skill

用于辅助安全审计、权限检查、凭据风险、认证流程和常见漏洞排查。它适合让 Agent 梳理敏感配置、检查依赖风险、分析鉴权逻辑或生成安全复核清单。使用时不能把工具输出直接当最终结论,涉及密钥、令牌、用户数据或生产系统时,应先确认最小权限、脱敏方式和操作边界。

总安装

220

周安装

9

GitHub Stars

37

下载量

71
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:mcp-oauth-setup(MCP OAuth 设置)
来源仓库:https://github.com/majesticlabs-dev/majestic-marketplace
仓库路径:skills/mcp-oauth-setup
安装命令:
npx skills add https://github.com/majesticlabs-dev/majestic-marketplace --skill mcp-oauth-setup
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/majesticlabs-dev/majestic-marketplace --skill mcp-oauth-setup

简介

用于辅助安全审计、权限检查和认证流程分析。mcp-oauth-setup 属于研究检索类 Skill,可作为该场景下的辅助能力补充。

  • 它能梳理敏感配置、检查依赖风险或生成安全复核清单。
  • 使用时不能把工具输出直接当最终结论,需人工复核。
  • 涉及密钥、令牌或用户数据时应确认最小权限和操作边界。
  • 建议优先脱敏处理,避免在生产环境中直接操作凭据。

SKILL.md

MCP Server Authentication & OAuth Dynamic Client Registration

Implement flexible authentication for MCP (Model Context Protocol) server connections. For OAuth providers, auto-discover endpoints and dynamically register as a client — the user just provides the MCP server URL and clicks "Connect." For bearer/API key providers, support both admin-shared and per-agent credentials so different agents can authenticate with different accounts.

When to Use

  • Building an admin UI for managing MCP server connections
  • Integrating with third-party MCP providers (Linear, Sentry, Granola, Render, etc.)
  • Implementing the MCP Streamable HTTP transport with authenticated tool sync
  • Adding per-agent credential support so each agent can use its own account
  • Adding OAuth to an existing MCP connector/server management system

Core Standards

The OAuth implementation relies on three RFCs:

  1. RFC 8414 - OAuth Authorization Server Metadata Discovery via .well-known/oauth-authorization-server
  2. RFC 7591 - Dynamic Client Registration at the provider's registration endpoint
  3. RFC 7636 - PKCE (S256) for authorization code security

Not all MCP servers use OAuth. Some (e.g., Render) use bearer tokens with API keys and handle account/workspace selection at the MCP protocol level. The credential system must be auth-type-agnostic.

Architecture Overview

Credential Mode (Orthogonal to Auth Type)

credential_mode applies to all auth types (bearer, api_key_header, oauth), not just OAuth. Different agents may need their own credentials for the same MCP server.

credential_mode = "shared"     → Admin provides one credential, all agents use it
credential_mode = "per_agent"  → Each agent has its own credential

OAuth Flow

Admin clicks "Connect"
    |
    v
Discover OAuth metadata (RFC 8414)
    |  GET /.well-known/oauth-authorization-server
    v
Register as OAuth client (RFC 7591)
    |  POST /oauth/register
    v
Redirect to provider consent screen
    |  GET /oauth/authorize?client_id=...&code_challenge=...
    v
Provider redirects back with code
    |  GET /callback?code=...&state=...
    v
Exchange code for tokens
    |  POST /oauth/token
    v
Store tokens, sync tools

Implementation Steps

1. Database Schema

Two tables: MCP server configuration (OAuth metadata + shared tokens) and per-agent credentials (any auth type).

See: references/schema.md

Key decisions:

  • Encrypt all secrets at rest (encrypts:oauth_client_id, etc.)
  • Store both shared tokens and per-agent tokens (join table)
  • credential_mode ("shared" or "per_agent") applies to ALL auth types
  • Store discovered_tools as JSON array
  • AgentMcpConnection.access_token stores OAuth tokens, bearer tokens, or API keys

2. OAuth Discovery and Registration

Three model methods on the MCP server record.

See: references/oauth_flow.md

  • Discovery (discover_oauth_metadata!): Derive .well-known/oauth-authorization-server URL, parse JSON response, skip if already configured, handle 404 gracefully (not all servers support RFC 8414)
  • Registration (register_oauth_client!): POST to registration endpoint, store client_id and client_secret, skip if already present
  • Combined (discover_and_register_oauth!): Run discovery then registration in sequence

3. Authorization Controller

Create an OAuth controller with authorize and callback actions.

See: references/oauth_flow.md

Critical pitfalls:

Turbo Drive cross-origin redirects: redirect_to with an external URL is silently swallowed by Turbo Drive — browser stays on current page. Use HTML with <meta http-equiv="refresh" content="0;url=..."> for the external redirect instead.

State parameter: Use a signed, expiring message (Rails message_verifier) with connector ID, PKCE code verifier, optional agent ID, and timestamp. Set 10-minute expiry.

String keys from message verifier: After verifying the state token, payload uses string keys not symbol keys. Use payload["connector_id"], not payload[:connector_id].

PKCE (S256): Generate a random code_verifier, compute code_challenge as URL-safe Base64 of SHA-256 digest with no padding.

Error redirects: When agent_id is present in state, redirect errors to the agent edit page, not the connectors index.

Auto-sync on first agent connection: For per-agent OAuth, when the callback stores the first per-agent token, auto-sync tools using that agent's token if tools haven't been discovered yet.

4. Routes

resources :connectors do
  member do
    get "oauth/authorize", to: "mcp_oauth#authorize", as: :mcp_oauth_authorize
  end
end
get "mcp_oauth/callback", to: "mcp_oauth#callback", as: :mcp_oauth_callback

Route helper naming: A member route mcp_oauth_authorize on resources:connectors generates mcp_oauth_authorize_connector_path(connector) — resource name comes last. Common source of NoMethodError.

5. Token Management

See: references/oauth_flow.md (ensure_token_fresh! pattern)

  • Check expiry with 5-minute buffer (token_expires_at < 5.minutes.from_now)
  • Use with_lock for thread-safe updates on shared tokens
  • Return appropriate token based on credential mode
  • Bearer/API key per-agent tokens are static (no refresh needed)

6. MCP Tool Sync (Streamable HTTP Protocol)

See: references/tool_sync.md

Two-step handshake:

  1. Send initialize JSON-RPC request → get Mcp-Session-Id header
  2. Send tools/list with session ID header

Critical details:

  • Set Accept: application/json, text/event-stream — some servers return 406 without this
  • Some servers return SSE format — parse both formats
  • sync_tools! must accept agent: parameter for per-agent auth
  • Some servers (e.g., Render) allow unauthenticated tool listing

7. UI Considerations

See: references/ui_patterns.md

Connector form:

  • credential_mode radio applies to ALL auth types
  • Hide admin token input when per-agent is selected for bearer/API key
  • Show OAuth fields only for OAuth auth type
  • Use Stimulus controller to toggle visibility based on both auth_type AND credential_mode

Agent edit form — three states for per-agent connectors:

  1. Per-agent OAuth, not connected → grayed card, "Connect" button
  2. Per-agent bearer/API key, not connected → inline password input
  3. Connected (any type) → tool checkboxes + "Token configured" badge

Verified MCP Providers

ProviderURLToolsAuthNotes
Linearhttps://mcp.linear.app/mcp45OAuthSSE response format
Sentryhttps://mcp.sentry.dev/mcp14OAuthStandard JSON
Granolahttps://mcp.granola.ai/mcp4OAuthStandard JSON
Renderhttps://mcp.render.com/mcp24Bearer tokenNo OAuth, per-agent API keys

Common Failure Modes

SymptomRoot CauseFix
Page stays on form, no redirectTurbo Drive swallows cross-origin 302Use HTML meta refresh instead of redirect_to
NoMethodError on route helperWrong helper name orderingMember route generates mcp_oauth_authorize_connector_path
payload[:connector_id] returns nilMessage verifier returns string keysUse payload["connector_id"]
406 from MCP serverMissing Accept headerAdd Accept: application/json, text/event-stream
400 "Mcp-Session-Id required"Skipped initialize handshakeSend initialize first, use returned session ID
JSON parse error on tool syncServer returns SSE formatDetect and parse both formats
Token exchange fails silentlyMissing code_verifierInclude PKCE verifier from signed state
OAuth discovery 404Server doesn't use OAuthUse bearer or API key auth instead
Per-agent connector shows no toolsAdmin can't sync without tokenTools auto-sync on first agent connection
Error redirect goes to wrong pageagent_id not checked in rescueRedirect to agent edit when agent_id present

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

37.71%
按下载量换算27

Claude

27.21%
按下载量换算19

Cursor

18.19%
按下载量换算13

Gemini CLI

8.49%
按下载量换算6

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

可疑

权限和风险

敏感数据

该 Skill 可能接触密钥、Token、环境变量或敏感配置,应进入高风险复核队列,默认不自动发布。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在 warning/failed 结果,不能写成本站确认安全。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills