Connect your agent.

Start with the prompt

This works with any client that can read a URL. The brief tells your agent where the endpoint lives, what consent looks like, and which read-only calls to make first. You approve the connection in your browser; the agent never sees your password. The prompt itself carries no secret, so pasting it into a chat costs nothing.

Keep it that way: never paste API keys, tokens or one-time codes into the conversation. Approvals happen on the AgentLinkOps authorization screen, and credentials stay in your client’s own secret storage.

Claude Code

Claude Code speaks streamable HTTP with OAuth, which is what the AgentLinkOps endpoint serves. Add the server with one line:

claude mcp add --transport http agentlinkops https://app.agentlinkops.com/mcp

tested 2026-09-12

The command form matches Claude Code 2.1.270, checked against its own help output. The server URL answers the production OAuth authorization envelope. The consent flow behind it passed a separate end-to-end run on September 9, 2026 against the staging deployment, which runs the same authorization code: consent, token exchange, refresh, downscope and revocation. One thing this page did not run: a whole session inside Claude Code against production, because that needs your browser login.

After the server is added, Claude Code opens the browser for sign-in and the consent screen. Approve the workspace connection, then ask your agent to verify with two read-only calls:

get_workspace({})
list_projects({ "limit": 20 })

Both need only projects:read. They create no monitors and start no checks.

Codex

The Codex CLI manages remote HTTP servers with its own subcommands:

codex mcp add agentlinkops --url https://app.agentlinkops.com/mcp

command form tested 2026-09-12

Checked against codex-cli 0.154.0 help output. The connection itself was not completed in this pass. When you connect, use the login subcommand to run the OAuth approval in your browser:

codex mcp login agentlinkops

The config file form is the documented equivalent of the add command, shown here for reference:

# ~/.codex/config.toml
[mcp_servers.agentlinkops]
url = "https://app.agentlinkops.com/mcp"

config form not tested here

Cursor

Cursor keeps remote servers in a JSON file in your home directory:

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "agentlinkops": {
      "url": "https://app.agentlinkops.com/mcp"
    }
  }
}

not tested

Cursor was not available in the environment that checked this page on September 12, 2026, so this snippet carries no test claim. It follows Cursor’s documented HTTP server format; check it against Cursor’s current documentation before relying on it.

REST and scoped API keys

If your tool is not an MCP client, the same service answers a REST API:

https://app.agentlinkops.com/v1/

Create a scoped API key in your workspace under Agent access. It is displayed once and expires within the period you choose. Send it as a bearer header, and store it in your tool’s secret storage rather than a command line or a checked-in file. The /mcp endpoint accepts these keys as well, which suits clients that cannot complete an OAuth screen. What each scope permits is described in the access section of the guide.

Not supported

Labeled plainly so your agent does not guess:

• Using an MCP OAuth token as a REST credential. The token is bound to /mcp.
• Local stdio servers. The endpoint serves streamable HTTP only; there is no command to pipe through.
• The anonymous free check through your authenticated connection. It lives on its own endpoint, /free/mcp, with one read-only tool.
• The repo-local context tools (context_*). They belong to a repository checkout, and no cloud endpoint serves them.
• Any client that cannot reach the public internet. Every surface above is hosted.

Everything your connection can call is listed, with honest states, in the tool catalog.

How this page was checked

On September 12, 2026, from this repository’s environment:

• Production /mcp answered with its OAuth authorization envelope, and its protected-resource metadata pointed at the Clerk authorization server with the eight documented scopes.
• The public brief at agentlinkops.com/SKILL.md answered 200 as Markdown.
• The Claude Code one-liner matched Claude Code 2.1.270 help output.
• The Codex command matched codex-cli 0.154.0 help output.

The consent flow behind these commands passed a separate end-to-end run on September 9, 2026 against staging: metadata discovery, client registration, browser consent, PKCE with a wrong-verifier rejection, token exchange over the MCP SDK, downscope and revocation. Cursor was not installed in that environment, so its snippet carries no test claim.

Hosted access is invitation-only. Reading this page, the brief or the catalog creates no account and starts no monitoring.