MCP Server

The Control Plane MCP Server lets any AI tool with Model Context Protocol support manage Control Plane resources through natural-language conversations.

If your AI tool supports plugins — Claude Code, Codex, or Antigravity CLI — start with the Control Plane AI Plugin instead. The plugin auto-configures this MCP server and adds skills, agents, commands, and guardrails on top.

Endpoint

string

required

https://mcp.cpln.io/mcp

Claude Web, Claude Desktop and ChatGPT don’t need this URL. Their directory listings are fixed on core, so enter it by hand only for mk8s, full, or readonly.

Authentication

OAuth with per-org consent. Your AI tool walks you through it on first use — see your tool’s setup page below for the exact trigger.

Organizations

On the consent screen you select the orgs the assistant may use, and that selection holds for the connection:

With one org selected, the assistant uses it without asking.
With several, it asks which one when you haven’t said.
It can’t reach an org you didn’t select, including one you create later. To add one, disconnect and reconnect, then select it.

list_orgs shows the selected orgs with their GVCs and locations.

Where apps run

When you ask for an app or a database and don’t name a GVC:

If the assistant created a GVC for apps earlier, it goes there, and the assistant says so.
Otherwise the assistant asks whether it goes in one of the org’s GVCs or in a new one, even when the org has only one.
If the org has none, the assistant recommends one of the org’s enabled locations near most of your users and creates a GVC named default in the location you choose. Adding a location to it later runs every workload in it there too.

A GVC you name that doesn’t exist yet is created in the location you choose. A GVC in several locations keeps a separate copy of every database and volume in each one, so the assistant warns before an app keeps data there. The assistant never moves, re-places, or deletes what already exists without your explicit yes, and it replaces a workload it didn’t create only after you confirm.

Toolset profiles

The full tool catalog is larger than most AI tools can hold in context at once — and some clients cap how many tools they accept. So each connection selects a toolset profile with the toolsets query parameter. Pick one — core, mk8s, and full are nested (each is a superset of the one above), and readonly stands on its own:

Profile	What it includes
`readonly`	Investigation only. Metrics, logs, traces, workload events, replicas and deployments, one-call workload diagnosis, quotas, the audit trail, configuration inventory, and Terraform export, plus a generic read tool that works on every resource kind. It advertises no tool that creates, updates, deletes or executes a command, so an assistant on this profile cannot change a live resource.
`core` (default)	The everyday deploy-and-operate journey. Job tools do a whole task in one call: deploy an app, add a database, create or rotate a secret without its value passing through the chat, and diagnose, restart, roll back, promote, or open a workload to other workloads. Around them: GVCs, workloads, deployments, logs, image builds (from a repository, or from app files the assistant writes itself, so a chat can take an app from the user’s request to a running URL), custom domains, identities, policies, the template catalog, volume sets, Terraform export, metrics, traces, and the audit trail, plus a generic read/delete tool that works on every resource kind.
`mk8s`	Everything in `core`, plus `create_mk8s` and `update_mk8s`, which create and update managed Kubernetes (mk8s) clusters on every supported provider.
`full`	Everything in `mk8s`, plus the full administrative and fine-tuning surface: cloud-account registration (AWS, GCP, Azure, NATS), org IAM (users, groups, service accounts), network-bridge agents, identity network-resource bindings, IP sets, audit contexts, external logging, advanced workload tuning (load balancer, sidecar, rollout, security, retry), domain CORS, volume-set snapshots and shrinking, GVC location edits, and org settings.

Examples

https://mcp.cpln.io/mcp                    # core (default)
https://mcp.cpln.io/mcp?toolsets=readonly  # reads only, no writes
https://mcp.cpln.io/mcp?toolsets=mk8s      # core + BYOK Kubernetes
https://mcp.cpln.io/mcp?toolsets=full      # everything

An unknown value returns a 400 listing the accepted names (readonly, core, mk8s, full). So does combining readonly with another profile — it withholds the write tools the others add, so the two cannot be mixed. The AI Plugin also adds skills=plugin, which leaves get_cpln_skill out because the plugin loads the Control Plane skills itself. Leave it off a connection that has no plugin.

Which profile should I use? full is the complete catalog: it’s what the AI Plugin configures for Claude Code, while it uses core for Codex and Antigravity CLI. Pick core (the default) for a lighter, focused catalog if your client caps tool counts or loads every tool definition up front; it still covers the whole build-and-operate workflow (it can read and delete any resource kind generically). Use mk8s when you manage BYOK Kubernetes clusters. Use readonly to point an assistant at production for performance work, incident triage, or an audit without the possibility of an unintended change.

The profile is fixed for the life of a connection — an AI agent can’t expand it mid-session. If your assistant says a Control Plane tool isn’t available, reconnect with the profile that includes it.

Compatible tools

Claude Code

Claude Code CLI

Claude Desktop

Claude apps for Mac & Windows

Claude Web

claude.ai in the browser

ChatGPT

ChatGPT web, desktop & mobile

Codex

OpenAI Codex

Cursor

Cursor IDE

VS Code

Visual Studio Code

Antigravity

Google Antigravity (CLI or IDE)

Amp

Sourcegraph Amp

OpenCode

OpenCode terminal assistant

Don’t see your tool listed? Any client that supports remote MCP servers can connect — point it at https://mcp.cpln.io/mcp and let it handle OAuth.

Next steps

Tools Reference

Every tool the server exposes, grouped by resource, with toolset profiles

Usage Examples

Practical prompts for common Control Plane workflows

AI Plugin

Recommended path for Claude Code, Codex, Antigravity CLI

Access Control

Scope what an AI tool can do in your org with policies and least privilege

Getting Started

MCP Clients

Endpoint

Authentication

Organizations

Where apps run

Toolset profiles

Compatible tools

Claude Code

Claude Desktop

Claude Web

ChatGPT

Codex

Cursor

VS Code

Antigravity

Amp

OpenCode

Next steps

Tools Reference

Usage Examples

AI Plugin

Access Control

​Endpoint

​Authentication

​Organizations

​Where apps run

​Toolset profiles

​Compatible tools

Claude Code

Claude Desktop

Claude Web

ChatGPT

Codex

Cursor

VS Code

Antigravity

Amp

OpenCode

​Next steps

Tools Reference

Usage Examples

AI Plugin

Access Control

Endpoint

Authentication

Organizations

Where apps run

Toolset profiles

Compatible tools

Next steps