MCP Integration

Muster ships with a built-in MCP (Model Context Protocol) server that lets AI assistants like Claude Desktop and Claude Code interact with your deploy infrastructure. Check status, trigger deploys, view logs, and roll back — all through natural language.

Setup

The MCP server is provided by the muster-mcp binary, which is installed alongside muster. Add it to your AI assistant's MCP configuration:

json

{
  "mcpServers": {
    "muster": {
      "command": "muster-mcp",
      "args": ["--project", "/path/to/your/project"]
    }
  }
}

Available MCP tools

The muster MCP server exposes these tools to AI assistants:

Setup & scaffolding

Name	Type	Description
`muster_scan_project`	tool	toolDetect project files (Dockerfile, compose, package.json, etc.) and suggest services.
`muster_init_project`	tool	toolCreate deploy.json, hook directories, and .gitignore entries.
`muster_write_hook`	tool	toolWrite a hook script (deploy.sh, health.sh, rollback.sh, logs.sh, cleanup.sh). Auto-creates dirs.
`muster_read_hook`	tool	toolRead an existing hook script. Shows available hooks if requested one is missing.

Operations

Name	Type	Description
`muster_list_services`	tool	toolList services with health type and hook availability.
`muster_status`	tool	toolHealth check all services (returns HEALTHY/UNHEALTHY per service).
`muster_deploy`	tool	toolDeploy all or a specific service, then verify health.
`muster_rollback`	tool	toolRollback a specific service.
`muster_logs`	tool	toolGet recent deploy log lines for a service.
`muster_cleanup`	tool	toolRun cleanup hooks and delete old log files.
`muster_config`	tool	toolReturn the full deploy.json / muster.json contents.

Credentials

Name	Type	Description
`muster_set_credential`	tool	toolStore a credential key-value pair in the MCP session (never written to disk).
`muster_list_credentials`	tool	toolList stored credential keys for the current session.
`muster_clear_credentials`	tool	toolWipe all in-memory credentials.

Warning

All tools require a project_dir parameter. The MCP server runs locally over stdio and requires jq. Credentials stored via muster_set_credential are kept in-memory only — they are never written to disk.

Example workflow

Once configured, you can interact with muster through natural language in Claude:

text

You: "What's the current deploy status?"
Claude: (calls muster_status) "All 3 services are healthy..."

You: "Deploy the API service with a dry run first"
Claude: (calls muster_deploy --service api --dry-run) "Dry run complete..."

You: "Looks good, deploy it for real"
Claude: (calls muster_deploy --service api) "Deploy successful..."

Tip

The MCP integration works with any AI assistant that supports the Model Context Protocol, not just Claude. Check your AI tool's documentation for MCP configuration.

Skills