Add an xAI preset: images, videos and speech with Grok Imagine

[nasser1941]

What

lcode mcp add xai first asks for an xAI API key: typed hidden, or taken from XAI_API_KEY if it's set. It then runs lcode-mcp-xai, a new MCP server shipped with lcode for xAI's Grok Imagine and text to speech APIs.

Tool	API
generate_image	POST /v1/images/generations: 1–10 images, aspect ratio, 1k/2k, model grok-imagine-image-2.0 (or -quality, or the basic one), as base64
edit_image	POST /v1/images/edits, with the local image as a data URL
generate_video	POST /v1/videos/generations, then polls GET /v1/videos/{id}: 1–15 s, up to 1080p, optionally from a local start image. Waits up to 10 minutes and downloads the MP4
video_status	Picks up a video that took longer than that
text_to_speech	POST /v1/tts: 26 voices, 20 languages, MP3 or WAV

• Where files go: generated/ in the project, or a path the model is asked for. An extension is added when missing, and nothing is overwritten.
• Errors: xAI's own messages reach the model, such as a moderation refusal, a bad key (with a hint about XAI_API_KEY) or a rate or credit limit.
• Music: xAI has no music generation API, so there's no music tool.
• Asking first: lcode asks before every call, since each is billed to the user's xAI account. "allow" in mcp.json turns that off per tool.
• Docs: a section in the MCP guide (tools, costs, privacy), a catalog row, and the README.

Tests

• tests/test_xai.py (6 tests) uses a fake xAI API. It covers:
  • the request bodies of every tool, and the saved files (names, extensions, no overwriting);
  • edits sending a data URL;
  • video polling, the request id handed back after the wait, and video_status;
  • TTS JSON with base64 audio;
  • moderation and bad-key errors;
  • the preset asking for the key.
• Real run with Qwen3.6 35B through Ollama against the fake API, with lcode's stdio MCP transport. The request was "make me an image of a red fox saved as assets/fox.png, a voice line in the Leo voice, and a 4 second video of the fox running starting from that image". The model called generate_image, text_to_speech and generate_video with the image as the start frame, and all three files were saved.
• Not run against the real xAI API: that needs a key and is billed.

🤖 Generated with Claude Code