Stability AI MCP server providing image generation, editing, upscaling, ControlNet, and 3D generation for Claude Code
The server defines 15 tools with explicit input schemas, descriptions, and parameter definitions visible in the source code. Naming follows verb_noun patterns consistently (generate_image, erase_object, inpaint, etc.), which is excellent. Descriptions are present for all tools and most parameters. However, several critical gaps reduce the score: (1) Output schemas are NOT documented, only inputs are defined. The source shows tool registration but no return type documentation, forcing LLMs to infer results. (2) Descriptions vary in quality; some are detailed (generate_image at 238 chars) but many are brief (check_balance has a 51-char description that barely meets minimum). (3) Error handling is not visible in the provided source, no evidence of recovery guidance, actionable error messages, or error classification. (4) Parameter constraints are inconsistent, seed parameters use JSON Schema min/max correctly, but several tools lack detailed format guidance in descriptions (e.g., file path parameters do not specify expected format). (5) Some parameters lack descriptions entirely or are too terse (e.g., 'output_format' in erase_object has no description, only enum). (6) Security: image_path parameters accept arbitrary filesystem paths with no visible validation or path-traversal protection in the provided source.
Check your Stability AI account credit balance.
Generate images from a sketch using ControlNet. The sketch guides the composition and structure. Trigger: 'stability sketch'.
Generate images guided by image structure/layout using ControlNet. Preserves composition while changing content. Trigger: 'stability structure'.
Generate images with a specific style by conditioning on a style reference image using ControlNet. Trigger: 'stability style'.
Erase objects from an image using Stability AI. Provide an image and optionally a mask to indicate what to erase. Trigger: 'stability erase'.
Generate a 3D model from an image or text description. Returns glTF or glb format 3D mesh. Trigger: 'stability 3d', 'stability mesh'.
No output schemas documented for any tool. LLMs cannot infer result structure, required fields for chaining, or pagination. This violates pattern:tool and pattern:response-shaper fundamentally.
File path parameters (image_path, mask_path, save_path) accept arbitrary strings with no visible validation or path-traversal protection. The source shows fs.writeFileSync() and path operations without sanitization. This enables directory traversal attacks if untrusted input is passed.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 54 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Generate images using Stability AI's Stable Diffusion models. Supports text-to-image and image-to-image with multiple models (ultra, core, sd3.5). Trigger: 'stability generate', 'stability image', or 'stable diffusion'.
Fill in masked areas of an image with new AI-generated content based on a text prompt. Trigger: 'stability inpaint'.
Extend an image's boundaries in any direction with AI-generated content. Trigger: 'stability outpaint', 'stability extend'.
Remove the background from an image, leaving a transparent background. Trigger: 'stability remove background', 'stability remove bg'.
Replace the background of an image with AI-generated content and optionally relight. Trigger: 'stability replace background'.
Find an object in an image by text description and recolor it. No mask needed. Trigger: 'stability recolor'.
Find an object in an image by text description and replace it with something else. No mask needed. Trigger: 'stability replace'.
Transfer the style from one image to another. Combines structure from one image with style from another. Trigger: 'stability transfer style'.
Upscale an image using Stability AI's upscaling models. Supports fast, conservative, and creative modes. Trigger: 'stability upscale', 'stability enhance'.
Descriptions for several tools are under 100 characters and lack actionable context. E.g., check_balance (51 chars) gives no guidance on when to call it.
Several parameters lack descriptions entirely or are minimal. E.g., 'output_format' in erase_object, remove_background, and others has only an enum, no description explaining why the LLM would choose one format over another.
No error handling logic visible in the provided source. No evidence of recovery guidance, actionable error messages (e.g., 'Invalid seed: must be 0-4294967294'), error classification (retryable vs fatal), or remediation steps. Per pattern:recovery-guide, errors must tell the LLM what to do next.
Image file parameters (image_path, mask_path, style_image_path) do not document expected formats, encodings, size limits, or supported MIME types. LLMs will pass any string. Should specify: 'Absolute path to a PNG/JPEG/WebP file under 10 MB'.
Parameter 'save_path' is optional in most tools but its default behavior is not documented. The source shows getAutoSavePath() generates auto-paths, but this behavior is not mentioned in parameter descriptions. LLMs will not know where files end up unless told explicitly.
No tool annotations visible (readOnlyHint, destructiveHint, idempotentHint). All 14 image generation/editing tools WRITE files and are destructive. Adding destructiveHint: true would signal to LLMs that these operations cannot be undone and require confirmation. Per current spec (2026-07-28), tool annotations are now standard.
Naming clarity: 'control_sketch', 'control_structure', 'control_style' share the 'control_' prefix but their distinctions are only clear from descriptions.