Raily MCP Server

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:

claude plugin marketplace add railyai/raily-mcp
claude plugin install raily@raily

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: partners@railyai.com

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:

cursor-agent mcp enable raily
cursor-agent mcp login raily
cursor-agent mcp list-tools raily

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

codex mcp add raily --url https://railyai.com/mcp
codex mcp login raily
codex mcp list

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:

grok mcp add --transport http raily https://railyai.com/mcp
grok mcp doctor raily --json

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:

{
  "mcpServers": {
    "raily": {
      "command": "npx",
      "args": ["-y", "mcp-remote@0.14.3", "https://railyai.com/mcp"]
    }
  }
}

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:

openclaw plugins install clawhub:@railyai/raily

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.

npm run catalog:sync   # refresh generated catalog from the live canonical registry mirror
npm run check          # offline manifest/version/catalog checks
npm run catalog:check  # fail if the server catalog has changed
mcp-publisher validate server.json
npm pack --dry-run

Release all manifest versions together. CHANGELOG.md records the bundle changes; the server remains independently deployed at the stable endpoint.