naba 

>_ mcp

naba as a Model Context Protocol server

naba mcp starts a stdio-based Model Context Protocol server that exposes naba's image pipeline as MCP tools, and its embedded skill tree as lazily-loaded MCP resources, to assistants like Claude Desktop and Cursor. It drives the same provider/selector/output pipeline the CLI uses — no generation logic is reimplemented.

Claude Desktop configuration

Add naba to your claude_desktop_config.json:

{
  "mcpServers": {
    "naba": {
      "command": "naba",
      "args": ["mcp"],
      "env": {
        "GEMINI_API_KEY": "<your-key>",
        "OPENROUTER_API_KEY": "<your-key>",
        "AWS_BEARER_TOKEN_BEDROCK": "<your-token>",
        "NABA_OUTPUT_DIR": "/path/to/output"
      }
    }
  }
}

Only the key(s) for the provider(s) you use are required. Verify the server responds to the MCP initialize handshake:

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0.1.0"}}}' | naba mcp

NABA_OUTPUT_DIR

MCP tools do not use the CLI's -o/CWD path. They write images through the MCP output-dir resolution: NABA_OUTPUT_DIR env > config default_output_dir > the XDG default ~/.local/share/naba/images. Setting NABA_OUTPUT_DIR is recommended so you know where generated images land. Errors surface as tool-level error results (isError: true), never a process exit.

Tools

The server registers exactly 8 tools:

Tool Required params Optional params
generate_image prompt style, variations[], count (1–8), seed, aspect, resolution, quality
edit_image prompt, file aspect, resolution, quality
restore_image file prompt, aspect, resolution, quality
generate_icon prompt sizes[], style, background, corners, format, quality
generate_pattern prompt style, colors, density, size, repeat, aspect, resolution, quality
generate_story prompt steps (2–8), style, transition, layout, aspect, resolution, quality
generate_diagram prompt type, style, layout, complexity, colors, aspect, resolution, quality
list_images — limit (default 20)

The generative tools accept the shared aspect, resolution, and quality params matching the CLI; generate_icon takes quality but no aspect/resolution (icon sizes are canvas pixels, a separate concept). With count/steps/sizes > 1 the same imageConfig applies to every image in the call. list_images is MCP-only — it lists recently generated naba-* images in the output directory, newest first.

Lazy-loading skills as resources

Beyond the generated file:// image links, the server exposes naba's embedded skill tree as MCP resources under a skill:// URI scheme — the lazy-loading pattern:

  • resources/list enumerates the embedded skill tree cheaply. Per embedded skill <name> it emits a compact skill://<name> index resource plus one skill://<name>/<rel> resource per file (SKILL.md, commands/*.md, README.md). Listing carries URIs and metadata only — never file bodies — so discovery is cheap.
  • resources/read serves content on demand:
  • skill://<name>/<rel> returns that embedded file as text (MIME by extension — text/markdown for .md).
  • skill://<name> returns a generated markdown index listing every file's read URI.
  • file://<path> returns a generated image by path (the reserved file:///{path} resource template).

A client discovers skills up front and pulls full instruction content only when it needs it, so tool schemas and the resource listing both stay lean.

  • Skills page — the same skill tree as a Claude Code slash command.
  • config — provider keys and default_output_dir.