Shells and terminals
Exec, debug containers, node shells, an in-cluster shell and your own local terminal — with the pods PodSteer creates cleaned up after.
MCP server and coding agents · New in v0.3.0
A read-only Kubernetes MCP server for your coding agent
The PodSteer binary is also a Model Context Protocol server. Run podsteer mcp from Claude Code or any MCP-capable agent and it can read your clusters — pods, logs, events, manifests, the assessment, RBAC — with the kubeconfig you already have. It serves over stdio, opens no port and offers no tool that writes.
Why read-only
Every write in PodSteer's interface is guarded by something a person reads: the type-the-name gate on production, the drain preview, the bulk review. An agent cannot be shown one of those. The honest options were a write with no confirmation or a confirmation nobody sees, so the server offers neither.
The nineteen tools
- Clusters, namespaces and kinds: list_clusters, list_namespaces, list_kinds.
- Objects: list_pods, list_workloads, list_nodes, list_resources, get_manifest, describe_resource.
- Diagnosis: assess_cluster, cluster_findings, assess_pod, get_logs, list_events, dependency_map, namespace_inventory.
- Access: rbac_subject_rules, rbac_can_i, rbac_inspect_role.
- cluster_findings returns only the ranked findings, each with the objects behind it, and can be filtered by severity and namespace; assess_cluster adds capacity and the node and pod summaries.
Answers an agent can trust
A refused read comes back as a refusal, not an empty list — an agent handed an empty array reports that the cluster has no such objects, with complete confidence. A truncated list states its true total, and a Secret's values are replaced with their size before anything is serialised.
Or start the agent from PodSteer
The local terminal can start a coding agent you already have — Claude Code, Codex, Gemini CLI or Copilot CLI — pointed at the cluster in the tab, without touching your kubeconfig's current context.
Be honest
What it does not do.
- No write tools, no exec and no port-forward, by design.
- An agent started from the local terminal runs with your credentials. Asking it to stay read-only is a request in its prompt, not a restriction PodSteer can enforce.
- PodSteer installs no agent; it starts the ones it finds on your PATH.
Read next
Cluster assessment
Ranked findings instead of a pod list: crash loops, OOM kills, pods that will never schedule, sizing problems — each with what to do about it.
Security and RBAC
Privileges a spec takes, vulnerabilities your scanner already found, and what an account can actually do — answered by the API server.