Skip to content

Coding agents and seylo-mcp

seylo-mcp is Seylo’s MCP server: one HTTPS endpoint that coding agents such as Claude Code use. With it, an agent can find out what is running in a namespace, check a deployment before applying it, deploy a pre-built image, read logs and metrics, and work out why a workload is failing. It does this through Seylo, so it never touches your cluster directly.

This page covers how it works and where its limits are. To set it up:

  1. An admin opens a namespace to coding agents. Nothing is open by default.
  2. The admin issues a deploy token for that namespace.
  3. Someone connects their client with the token.

Two agents, and only one of them can reach your cluster

Section titled “Two agents, and only one of them can reach your cluster”

There are two agents involved, and they hold different things.

The coding agent (Claude Code, for example) Seylo’s in-cluster agent
Runs On someone’s laptop or in their tooling Inside your cluster, as one Deployment
Holds A deploy token: a capability for one namespace on one cluster Its own ServiceAccount and the RBAC you installed it with
Talks to https://portal.seylo.in/api/mcp, over HTTPS Seylo, over an outbound WebSocket it opens itself
Can reach the Kubernetes API No Yes, within its RBAC

The coding agent asks for things and Seylo decides. If the request is allowed, Seylo sends it to the in-cluster agent, which applies it. The coding agent gets no kubeconfig, cloud credential or registry credential, so none of those can end up in a model’s context or a chat transcript.

All of Seylo’s usual guarantees still apply. Seylo holds no cloud credentials for agent-managed clusters, and there is no inbound connection to your cluster. See the trust model.

What it can do depends on the capabilities ticked when its token was issued. The default is read-only.

  • Look: see its namespace, what is deployed there, recent operations, pod logs, ConfigMap values, and Secret names and key names.
  • Observe: query metrics, logs and traces, and read dashboards, alerts and silences. Every query is limited to its one namespace. See the tool reference.
  • Check: generate manifests from a small spec, then test them against the policy engine and a server-side dry run in your cluster. Nothing changes.
  • Change, if the capability was ticked: deploy a pre-built image, roll a workload back, and add keys to a ConfigMap.
  • Propose, if the capability was ticked: save a dashboard as a draft, or ask for a Deployment’s metrics to be scraped. A person decides on each one. See proposals and approval.
  • Ask for a Secret: record that the application needs a Secret, with its name and key names. A person creates it. The agent never supplies the value and never sees it.

None of these depends on a setting. There is no tool that would do them.

  • Read a Secret value. No tool returns one, and error messages don’t either. See how secrets are handled.
  • Reach a second namespace or cluster. No tool takes a namespace or cluster argument. If an agent sends one anyway, the call is refused and the refusal names the argument.
  • Approve anything, including its own change. The MCP server has no approval tool.
  • Widen its own token, issue a token, open a namespace, or change a namespace’s quota.
  • Exec into or attach to a container. No MCP tool does either.
  • Delete anything. No tool deletes. ConfigMap writes only add or update keys.
  • Choose its own security settings. The deploy spec has no fields for privileged mode, host networking, service accounts or volumes. Seylo sets those.
  • Override a reconciler. deploy, rollback, put_config and scrape_target_propose check who owns every object they would change. If Argo CD, Flux, Helm or an operator owns one, the call is refused and the refusal names the owner. A coding-agent token can’t acknowledge another owner; only a person can, in Seylo, on the record.

The server offers MCP tools only. Resources, prompts, sampling and roots aren’t implemented, so a client that asks for them gets “method not found”.

How a change flows depends on whether it applies directly or goes to a person as a proposal.

  1. Orient. The agent calls list_targets to learn its boundary, then describe_target to see what is running there.
  2. Check. validate generates the manifests from a spec, evaluates them against the policy catalogue, and runs a server-side dry run in your cluster. It changes nothing and can be called as often as the agent likes.
  3. Deploy. deploy repeats the whole check rather than trusting an earlier validate. Nothing is applied if any of these is true:
    • a policy finding is blocking;
    • the cluster’s dry run rejects the manifests;
    • the namespace has no ResourceQuota;
    • Argo CD, Flux, Helm or an operator owns an object the deploy would write;
    • the token has used up its daily cap.
  4. Apply. The in-cluster agent applies the manifests. Seylo records an ordinary operation, attributed to the person who issued the token, with the token’s label in the summary. The generated spec, the policy result and the dry-run result are stored with it.
  5. Watch. Applied doesn’t mean running yet. The agent polls operation_status until the operation succeeds or fails, then uses diagnose and logs if it failed. rollback undoes the change.

Each deploy and rollback, and each deploy refused by policy, by the dry run or because another controller owns its target, is emailed to the person who issued the token.

dashboard_propose and scrape_target_propose apply nothing. A dashboard becomes a draft that a releaser publishes or discards. A scrape change goes into the approval queue like any other manifest edit, and nothing is applied until an approver approves it. See proposals and approval.

Every tool call counts against a per-token limit, in two separate buckets: 60 calls a minute for tools that change nothing, and 10 calls a minute for deploy, rollback, put_config, request_secret, dashboard_propose and scrape_target_propose. The daily deploy cap applies on top. A runaway loop is refused with a message saying how long to wait, and a busy read loop can’t block a deploy. See limits.

Seylo records every call, including refusals. The record says which token made the call, which tool it used and whether it succeeded. Each observability call also gets an audit row with the query text, redacted, and the size of the result. The result itself isn’t recorded, because it can contain log lines. Pod logs pass through to the agent and are never stored by Seylo.

If changes are blocked for your organisation - it is suspended, or a free trial Seylo granted it has ended - the tools that change something are refused. These are deploy, rollback, put_config, request_secret, dashboard_propose and scrape_target_propose. Every read tool keeps working.