@kunobi/mcp

MCP bridge to Kunobi, a desktop platform management IDE. AI assistants manage Kubernetes, FluxCD, ArgoCD, and Helm while users maintain real-time visual oversight.

What is Kunobi?

Kunobi is a desktop IDE for platform engineering — built with Rust and React, no Electron.

• Real-time cluster visibility with resource browser, YAML editor, and embedded terminal

• Native FluxCD, ArgoCD, and Helm support

• Built-in MCP server for AI assistants (Claude Code, Cursor, Windsurf, Codex CLI, Gemini CLI, GitHub Copilot CLI)

• Available on macOS, Windows, and Linux

• No account required, no cloud dependency

Related MCP server: veriloop

Setup

Enable MCP in Kunobi under Settings > AI & MCP, then install:

Automatic (recommended)

Register with all your AI clients in one step:

npx @kunobi/mcp --install

This interactively detects your installed AI clients and registers the server with them. Supported clients:

• Claude Code — project or user scope

• Claude Desktop — user scope

• Cursor — project or user scope

• Windsurf — project or user scope

• Codex CLI — project or user scope

• Gemini CLI — project or user scope

• GitHub Copilot CLI — user scope

To remove the server from all clients:

npx @kunobi/mcp --uninstall

Updating

install pins your clients to an exact version (e.g. @kunobi/mcp@1.2.3) so each launch runs from the npm cache instead of resolving latest from the registry on every spawn — startup stays fast and can't stall on the network. When a newer version is published, the server tells you on connect; apply it with:

npx @kunobi/mcp upgrade

This re-pins every client where Kunobi is registered to the newest version and warms the cache. Restart your AI client to load it (the new version takes effect on the next start).

Installed before pinning existed? Pin your current version in place — no upgrade, no registry call:

npx @kunobi/mcp pin

If you ran an older build, the server also hints this on connect when it detects an unpinned config. Silence the hint with MCP_KUNOBI_NO_PIN_HINT=1.

Prefer always-latest? Opt back out — each launch will resolve the newest version again (slower startup, and can stall on the registry):

npx @kunobi/mcp unpin

Manual

If you prefer manual setup, add the following to your client's MCP config:

{
  "mcpServers": {
    "kunobi": {
      "command": "npx",
      "args": ["-y", "@kunobi/mcp"]
    }
  }
}

How it works

AI assistant <--stdio--> @kunobi/mcp <--HTTP--> Kunobi variants
                          │
                          ├── keeps one MCP connection per configured variant
                          ├── retries disconnected variants every 5s
                          └── manages one bundler per variant

The server keeps a persistent MCP connection for each configured Kunobi variant. When a variant is available, its tools are registered with a variant__ prefix (e.g., dev__list_clusters, stable__query_store). When a variant briefly drops, the hub keeps the last-known surface stable while it reconnects in the background. If the disconnect outlives the reconnect grace window, the stale registrations are removed.

Multi-variant support

Kunobi ships multiple release channels that run on different ports. This MCP server discovers all of them simultaneously:

These defaults are auto-generated into ~/.config/kunobi/mcp.json on first run. You can add custom variants or change ports — see Configuration.

Built-in tools

These are always available, even when no Kunobi instance is running:

• kunobi_status — reports all variant connection states, ports, and tool counts

• kunobi_launch — launches a Kunobi variant by name

• kunobi_refresh — forces an immediate reconnect attempt across all configured variants

• kunobi_call — stable entrypoint for variant tools (variant, tool, arguments)

Recommended calling pattern

Use the stable path below for the most reliable MCP client behavior:

1. Call kunobi_status

2. Read kunobi://tools to discover full downstream tool schemas and metadata

3. Execute via kunobi_call(variant, tool, arguments)

Dynamic tools

When a Kunobi variant is detected, its tools appear automatically with a variant prefix. For example, if dev and stable are both running:

• dev__app_info, dev__query_store, dev__list_stores, ...

• stable__app_info, stable__query_store, stable__list_stores, ...

Tools appear dynamically as variants start, and they are withdrawn only after a sustained disconnect — no MCP server restart needed.

These dynamic variant__tool entries are still supported, but some MCP clients may not refresh dynamic tool lists reliably. Use kunobi_call as the primary path when in doubt.

Resources

Resources exposed by Kunobi variants are proxied through automatically. When a variant connects, its resources become available to the AI client with variant-namespaced URIs to avoid collisions between multiple running variants.

The server also provides built-in resources:

• kunobi://status — JSON snapshot of all variant connection states, ports, and capability counts. Supports subscriptions — clients receive notifications/resources/updated whenever variants connect or disconnect.

• kunobi://tools — JSON discovery document listing each variant status plus full downstream tool, resource, and prompt metadata for kunobi_call.

Prompts

Prompts from Kunobi variants are registered with a variant__ prefix (e.g., dev__setup_cluster). They appear and disappear alongside their variant, just like tools.

Configuration

Config file

Variant-to-port mappings are stored in ~/.config/kunobi/mcp.json (auto-generated on first run with defaults). You can edit this file directly or use the CLI:

# List all configured variants and their connection status
kunobi-mcp list

# Add a custom variant
kunobi-mcp add juan 4200

# Remove a variant
kunobi-mcp remove juan

Environment variables

Priority: config file defaults → MCP_KUNOBI_VARIANTS env var (merges on top).

CLI

kunobi-mcp [command] [options]

Development

pnpm install
pnpm build
pnpm test

For local development against @kunobi/mcp-bundler:

pnpm dev:link    # link local bundler
pnpm dev:unlink  # revert to npm version

License

Apache-2.0