Control Plane MCP Server

<p align="center"> <a href="https://controlplane.com"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://cdn.jsdelivr.net/gh/controlplane-com/brand@main/logo/Control_Plane_logo_full_light.svg"> <img src="https://cdn.jsdelivr.net/gh/controlplane-com/brand@main/logo/Control_Plane_logo_full_dark.svg" alt="Control Plane" width="240"> </picture> </a> </p>

Control Plane AI Plugin

Run containerized workloads across AWS, GCP, Azure, OCI, and your own hardware under one API. It loads Control Plane's domain knowledge, production guardrails, and live MCP tools into Claude Code, Codex, and Antigravity CLI so your assistant can deploy, troubleshoot, secure, and migrate workloads with verified cpln commands.

Installation

Claude Code

/plugin marketplace add https://github.com/controlplane-com/ai-plugin.git
/plugin install cpln@controlplane
/reload-plugins

Update with /plugin marketplace update controlplane then /reload-plugins (third-party marketplaces don't auto-update unless you enable it in /plugin → Marketplaces).

Alternatively, install from Anthropic's community plugin marketplace, where every listing has passed Anthropic's review:

/plugin marketplace add anthropics/claude-plugins-community
/plugin install cpln@claude-community
/reload-plugins

Codex

codex plugin marketplace add https://github.com/controlplane-com/ai-plugin.git

Start Codex, open /plugins, and install cpln. Guardrail injection needs plugin hooks, which Codex gates off by default — enable them in ~/.codex/config.toml and restart:

[features]
plugins = true
plugin_hooks = true

Update with codex plugin marketplace upgrade controlplane, then restart Codex.

Antigravity CLI

Install the plugin with Antigravity CLI (agy):

agy plugin install https://github.com/controlplane-com/ai-plugin/plugins/cpln

Generic MCP client

Point any other MCP client at the hosted server:

{
  "mcpServers": {
    "cpln": {
      "type": "http",
      "url": "https://mcp.cpln.io/mcp?toolsets=full"
    }
  }
}

Authentication

MCP uses OAuth 2.1 + PKCE. Sign in to let the assistant act on your Control Plane organizations — you choose which orgs it may operate on, and the token is scoped to those orgs and enforced server-side on every call. Sign in again to change the grant. Treat MCP access as production access to the orgs you grant. How to sign in:

• Claude Code — /mcp, select cpln, sign in (or claude mcp login cpln).
• Codex — codex mcp login cpln.
• Antigravity CLI — /mcp, select cpln, authenticate.

Environment variables

Optional — only for the cpln CLI workflows some skills generate (CI/CD, Terraform, Pulumi). See .env.example.

Variable	Purpose
CPLN_TOKEN	Service-account token for cpln CLI calls (sensitive).
CPLN_ORG	Default organization.
CPLN_GVC	Default GVC.
CPLN_PROFILE	Local cpln CLI profile.

Usage

Ask in natural language — the assistant routes to the right skill or agent:

• "Troubleshoot why my payments-api workload in production keeps restarting."
• "Put app.example.com in front of my web workload with auto-TLS."
• "Give my analytics workload credential-free read access to S3 bucket prod-event-logs — no IAM keys."
• "Provision a production Postgres with HA failover and S3 backups."
• "Convert this kustomization.yaml to Control Plane and apply it to staging after I confirm."

Two workflows also have slash commands in Claude Code — /cpln:troubleshoot WORKLOAD and /cpln:migrate-k8s FILE; in other clients, ask for the same workflows by name.

What's included

• Domain skills across CLI usage, access control, autoscaling, networking, observability, migration, templates, stateful storage, and security.
• Two guided agents: workload troubleshooting and Kubernetes / Compose / Helm migration.
• An always-on guardrail rule the assistant applies in every session.
• Pre-configured access to the hosted Control Plane MCP server.

Security

• MCP access is production access — scoped to the orgs you grant and your own RBAC.
• Destructive actions (deleting resources, shrinking/deleting volumes, replacing workloads, applying to production) require explicit confirmation.
• Secret values are never exposed through the MCP tools — secrets surface as metadata only.
• The plugin stores no logs, secrets, prompts, or telemetry; your AI client and model provider process prompts per their own policies.

Report vulnerabilities per SECURITY.md.

More

• Contributing: CONTRIBUTING.md
• Docs: docs.controlplane.com · Support: support@controlplane.com or Slack
• License: MIT — see LICENSE