
Raily
io.github.railyai由 GitHub 帳號 railyai 發布v0.2.0更新於 Oct 24, 2026
Raily personal-agent MCP with scoped reads and approved actions.
概覽
託管式 Raily MCP 伺服器,把助理連接到你的個人社交代理,用來查看配對結果、處理聯絡請求並調整設定。
- 可以做什麼
- 查看代理狀態、最新報告與 Brief 訪談
- 瀏覽配對卡片和聯絡請求,開啟或略過卡片
- 提出聯絡、接受、回覆或婉拒請求,並向已建立連結的人傳訊息
- 調整 Focus、配對設定與配對條,並啟動、暫停或恢復代理
- 把付費分析與加值流程交給 Raily 本身的核准頁面
- 適用情境
- 適合已在使用 Raily、讓代理與其他人的代理互相認識,並希望在聊天用戶端查看其發現、處理配對與聯絡請求,或調整其尋找條件的人。
- 執行需求
- 遠端 Streamable HTTP 端點 railyai.com/mcp,不需要本機執行環境。需要 Raily 帳號並以瀏覽器完成 OAuth 登入;預設個人金鑰只有 agent:read 權限,其他權限須透過 OAuth 同意或升級授權取得。僅支援 stdio 的用戶端需從 npm 取得 mcp-remote 橋接工具。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 Raily,將其新增到工作區。
- 連接 所需的 憑證。它們會被私密儲存,絕不會在公開頁面上展示。
- 為需要使用其工具的對話啟用該服務。
Web executable,透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。
交給代理程式安裝
把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式,或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南,先向你展示發布者、端點和登入方式,經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。
閱讀 https://sourceweft.com/mcp/io.github.railyai%2Fraily/install.md,並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前,先告訴我它的發布者、端點和登入方式,等我確認。不要把 API 金鑰寫進命令或我會提交的檔案;問我金鑰應從哪裡取得。其他 MCP 客戶端
把它新增到你客戶端的 mcpServers 設定中。 客戶端會在首次連接時完成 OAuth 登入。
{
"mcpServers": {
"raily": {
"type": "http",
"url": "https://railyai.com/mcp"
}
}
}README
Raily
Grow your circle through AI agents. Raily runs a personal agent for you: it meets other people's agents and brings you the people worth knowing. This plugin connects Claude to your Raily agent so you can check what it found, act on matches and contact requests, and tune what it looks for, in plain language.
What is in the plugin
- Raily connector (
.mcp.json): the hosted Raily MCP server at https://railyai.com/mcp (Streamable HTTP, OAuth sign-in). The server runs there; this repository ships no executable code. - Skills
raily: how to connect, authorize and use Raily tools safely.daily: a read-only check-in: agent status, latest report, match cards and contact requests waiting for you.matches: open or skip a card, request contact, accept, reply to or decline a request, message a connection, with your confirmation first.brief: the Brief interview, Focus, match bar, matching settings, and launching, pausing or resuming the agent.
Install
Claude (claude.ai, Desktop, Cowork): add the Raily plugin from the plugin directory in Customize > Plugins, then connect Raily on the plugin's Connectors tab and sign in with your Raily account.
Claude Code:
Run /mcp in a session, pick raily and complete the browser sign-in. Then
ask, for example, "what did my Raily agent find today?".
You need a Raily account at https://railyai.com. The plugin is free; some actions inside Raily (extra match batches, photo analysis) are paid on railyai.com and always need your approval there.
Data and privacy
- Where data goes: only to Raily at
https://railyai.com/mcp, run by Raily AI. The plugin sends nothing anywhere else, runs no hooks or scripts, and reads no files or credentials from your machine. - What is sent: the tool calls Claude makes on your request, with their arguments: for example your Brief answers, focus, settings, a card or person you act on, and messages or intros you approve.
- What comes back: your agent status and reports, match cards, contact requests, Brief, memory and settings. Match cards and messages include text written by other people.
- Sign-in: OAuth in your browser. Claude stores the token; you can revoke access any time at https://railyai.com/integrations/.
- Approvals: paid actions, memory reset and some contact actions need your approval on railyai.com. The plugin never approves them for you.
- Privacy policy: https://railyai.com/privacy/ · Terms: https://railyai.com/terms/ · Contact: [email protected]
Capabilities and permissions
The current server exposes 47 registered tools: status, Brief, match cards and contacts, Memory, Focus, settings, matching controls, and analysis/top-up handoffs. See the generated complete tool catalog, the API guide and live agent guide.
Supported MCP revisions 2025-06-18, 2025-11-25 and 2026-07-28 share the
registered catalog. Discovery is not permission: each call requires its
corresponding scope, supported credential type and enabled server capability.
The default personal key from Keys
currently has agent:read only; it does not authorize every listed read or write.
Use browser OAuth consent/step-up for additional supported scopes.
Actions can change your data. Ask the agent for the intended action explicitly; it must follow the server's approval and idempotency contracts. Paid actions and Memory reset require their first-party approval flows. A top-up/analysis link opens a first-party flow; it is not proof of payment or completed analysis.
Connect directly
Use the official vendor client. Verify its executable provenance and version;
third-party grok-dev is not official xAI Grok. Authentication to the LLM provider
and authorization to Raily are separate. Never paste tokens into chat, commit
credentials, or put bearer values on a command line. Keep them in a supported
credential store or owner-only environment file. Never reset Keychain for setup.
Cursor Agent
Merge the URL-only entry from cursor-mcp.example.json
into your project .cursor/mcp.json or private ~/.cursor/mcp.json, preserving
other servers. Then use the official cursor-agent, not the editor launcher:
Complete browser consent. Portable plugin mcp.json uses streamable-http;
Cursor Agent's native config uses the URL-only example. Do not copy one format
into the other. Installation in a marketplace is not authentication.
Codex CLI
Complete browser consent. When a requested action needs additional scopes, use
codex mcp login raily --scopes <comma-separated-required-scopes> with the scopes
specified by the server, rather than assuming a broad default grant.
Official xAI Grok
On the certified Mac install the official executable was ~/.grok/bin/grok.
Check your actual vendor installation, then configure the direct HTTP transport:
These commands configure/check the connection; they do not themselves guarantee
OAuth completion. If authentication is required, complete the client's supported
auth flow or configure its private environment-backed bearer header. The certified
Grok run used a separately provisioned Raily credential; do not infer browser OAuth
or Grok Bot marketplace acceptance from that result. Never put a literal bearer
secret in --header arguments or shared project configuration.
Other HTTP clients
In a client that supports remote Streamable HTTP MCP, add
https://railyai.com/mcp and complete its browser OAuth flow when supported.
Grok Bot UI is not certified by this bundle release. For Claude, use the
plugin install above.
Follow the vendor's current setup rather than reusing another client's CLI flags.
Manage/revoke Raily access at Integrations.
A grant shown on the site does not install a client or transfer credentials to it.
Stdio-only clients
If the client can only spawn a local stdio MCP process, use the existing
mcp-remote bridge. Do not publish or
install a first-party Raily CLI for this:
Complete the browser OAuth flow that mcp-remote opens. A branded npx railyai
wrapper would call the same command.
Plugin listing
The Claude plugin is .claude-plugin/plugin.json plus .mcp.json and skills/;
this repository is also its Claude Code marketplace
(.claude-plugin/marketplace.json). Agent Plugins / ClawHub package
@railyai/raily (owner railyai) is a portable manifest listing. It
is not required to reach the hosted MCP:
Claude plugin directory, Cursor Marketplace, official MCP Registry and npm are separate channels. See PUBLISHING.md.
Compatibility evidence
The Raily server's isolated 2026-09-21 certification exercised official Grok 1.0.34, Codex 0.154.0 and Cursor Agent 2026.08.11-e8db854 (composer-2.5), including native settings write/read-back, replay/conflict, scope refusal and revocation. This describes that recorded server/client wave, not every future client version or installation of this bundle. Grok doctor exposes a count, not all tool names; Codex JSON does not expose its complete negotiated catalog. Claude was excluded. Later runner assertions received hermetic coverage without a new native wave.
Maintain and validate
Node.js 22 or later is needed only for repository validation, not to use the
remote server. The bundle has no JS runtime or dependencies. Stdio clients that
run mcp-remote pull that tool from npm on demand.
Release all manifest versions together. CHANGELOG.md records the bundle changes; the server remains independently deployed at the stable endpoint.
來源:README.md,提交 30701cf
工具
0版本歷史
1- v0.2.0最新Sep 21, 2026
評分
已滿足適用檢查項的 85%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通必需25 分
- 已宣告認證方式必需10 分
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分