Skip to content

Connect your coding agent

Connecting a coding agent takes two things: the seylo-mcp endpoint and a deploy token. The token is shown once, so have your client ready before you issue it.

What you need: an admin to issue the token, and a namespace that is already open to coding agents.

Endpoint https://portal.seylo.in/api/mcp
Transport MCP Streamable HTTP. One path, POST only, stateless
Authentication The header Authorization: Bearer <deploy token>
Protocol versions 2026-07-28, 2025-11-25 and 2025-06-18

The token screen shows the endpoint in its MCP endpoint field. Copy it from there.

A few things about the transport that clients notice:

  • No SSE stream. A GET to the endpoint answers 405. Long-running work is followed by polling, either with operation_status or with a task handle. See long-running work.
  • No session. Seylo issues no Mcp-Session-Id. Every request carries the token and is authenticated on its own.
  • Not from a browser. A request carrying a browser Origin other than Seylo’s own is refused. Coding agents send no Origin, which is the normal case.
  • Only the bearer token. Seylo has no OAuth flow for this endpoint, and a user’s sign-in session is not accepted.
  1. Open the dialog.

    Admin → Clusters, then Connect a coding agent on the Coding agents card. The button is disabled until some namespace is open to coding agents.

  2. Fill it in.

    • Label: how you will recognise this token later. Name the person or the machine, such as Priya's Claude Code. 1 to 80 characters.
    • Cluster and Namespace: only open namespaces are listed.
    • Capabilities: read-only by default. Tick only what this agent needs. Tick deploy, rollback and put_config only if the agent should change the cluster. The dialog marks those three “changes the cluster”. See capabilities.
    • Expires in (days) and Daily deploy cap: see tokens.
  3. Press Issue token.

  4. Copy the token now.

    This is the only time Seylo shows it. Seylo stores only a hash, so nobody can display it again, including a platform admin. If you lose it, revoke it and issue another.

The token screen also shows ready-made connection commands with your token already in them. Run one while the token is still on screen.

Run the command from the token screen. It has this shape:

Terminal window
claude mcp add --transport http seylo https://portal.seylo.in/api/mcp --header "Authorization: Bearer <deploy token>"

Add --scope user to make it available in every project rather than only the current one.

The token is stored in Claude Code’s MCP configuration. Treat that file as a secret, and don’t commit it to a repository.

Ask the agent to call list_targets. It should answer with exactly one target, your cluster and namespace, the capabilities the token carries, and the per-minute call limits:

{
"targets": [
{
"cluster": "acme-dev",
"namespace": "shop-dev",
"capabilities": ["list_targets", "describe_target", "operation_status", "logs", "observe"],
"dailyDeployCap": 20,
"issuedFor": "Priya's Claude Code"
}
],
"callLimits": {
"readCallsPerMinute": 60,
"writeCallsPerMinute": 10,
"writeTools": "deploy, rollback, put_config, request_secret, dashboard_propose, scrape_target_propose",
"note": "Per token, rolling one-minute windows, counted separately. A refused call counts. Writes are also bounded by dailyDeployCap."
},
"note": "This is the whole of your access. Other namespaces and clusters are not reachable with this token, and no tool here can change that."
}

A client sees only the tools its token can call. Tools for capabilities that weren’t ticked aren’t listed at all. A client may cache the tool list for up to five minutes.

If something fails, find the message on troubleshooting.