Release: v0.10.0

Fixed

• Auto-generated filenames now use the correct extension based on the Gemini API response MIME type (previously always .png)
• User-provided filenames without an extension now get one appended automatically
• MCP response MIME type now reflects the actual image format instead of being derived from the file extension
• Input image MIME type is now detected from the file extension instead of being hardcoded to image/jpeg

Changed

• MIME type and extension mappings are now centralized in a single utility module
• Unknown MIME types from the API are validated against a supported allowlist

Release: v0.9.0

Breaking Changes

• Minimum Node.js version raised from 20 to 22 (Node.js 20 EOL: April 2026)

Changes

• Update engines field in package.json to >=22
• Update CI matrix from Node.js 20.x to 22.x

Release: v0.8.2

Security

• Validate input image paths before reading: block path traversal (..), null byte injection, and resolve symlinks (#74)
• Enforce image file extension check on inputImagePath to prevent arbitrary file reads
• Sanitize user-supplied fileName before path construction
• Replace Math.random() with crypto.randomBytes() for file naming
• Pin GitHub Actions to SHA hashes (#67)

Improvements

• Better structured prompt generation for intent preservation, spatial clarity, and mood specificity (#73)
• Improve image-generation skill for LLM execution quality (#72)

Dependencies

• Bump path-to-regexp 8.3.0 → 8.4.0
• Bump brace-expansion, picomatch
• Update non-major dependencies

Release: v0.8.1

What's New

• Google Image Search grounding: When useGoogleSearch is enabled, the model now retrieves both web and image search results before generating images. This provides visual references for real-world subjects (landmarks, products, styles), improving generation accuracy.

Details

• Updated googleSearch tool configuration to include searchTypes: { webSearch, imageSearch }
• Added test coverage for tools parameter structure verification

Note

• Image Search grounding is only supported on gemini-3.1-flash-image-preview (Nano Banana 2). When using quality: "quality" (Nano Banana Pro), only web search grounding is effective.

Release: v0.8.0

Nano Banana 2 Support & Quality Presets

What's New

• Nano Banana 2: Default image generation model is now gemini-3.1-flash-image-preview (Nano Banana 2), delivering faster generation at lower cost
• Three quality presets: Control the speed/fidelity tradeoff via IMAGE_QUALITY environment variable or per-request quality parameter
  • fast (default) — Nano Banana 2, best for drafts and rapid iteration
  • balanced — Nano Banana 2 with enhanced thinking, better detail and coherence
  • quality — Nano Banana Pro (gemini-3-pro-image-preview), highest fidelity for final deliverables
• New aspect ratios: Added 1:4, 1:8, 4:1, 8:1 for ultra-tall and ultra-wide formats
• Updated image sizes: 1K, 2K, 4K options (removed 512px)

Improvements

• LLM-optimized tool descriptions for more accurate parameter usage by AI assistants
• Server-configured quality default is now respected — AI assistants only override when explicitly asked

Release: v0.7.0

Gemini 2.5 Flash Migration

Gemini 2.0 Flash is being deprecated at the end of March 2026. This release migrates the prompt generation model to Gemini 2.5 Flash with tuned parameters to maintain output quality.

What's Changed

• Model: Prompt generation now uses Gemini 2.5 Flash (previously 2.0 Flash)
• SDK: Updated @google/genai from v1.30.0 to v1.42.0
• Output quality: Tuned generation parameters (thinkingBudget, maxTokens, topP, topK) through iterative evaluation to produce concise, high-density prompts
• STYLE directives: Enhanced to encourage specific artistic references and camera specifications in generated prompts
• Docs: Updated README and SKILL.md to reflect the new model
• Security: Bumped ajv from 8.17.1 to 8.18.0 (CVE-2025-69873 ReDoS fix)

Release: v0.6.0

Agent Skill: Image Generation Prompt Best Practices

This release adds a standalone Agent Skill that teaches AI assistants how to write better image generation prompts — no MCP server or API key required.

If your AI tool already has built-in image generation (e.g., Cursor 2.4+), install just the skill:

npx mcp-image skills install --path ~/.cursor/skills

The skill covers the Subject-Context-Style framework, camera/lighting terminology, character consistency, compositional integration, and more. Works with any tool supporting the Agent Skills standard.

Other Changes

• Entry point refactored into thin router (index.ts) + server logic (server-main.ts)
• Added agent-skills and skills keywords to package metadata

Release: v0.5.6

What's Changed

Dependencies

• Bump @modelcontextprotocol/sdk from 1.25.2 to 1.26.0
  • Security fix for cross-client response data leak (GHSA-345p-7cg4-v4c7)
• Bump hono from 4.11.4 to 4.11.7
• Bump lodash from 4.17.21 to 4.17.23

Full Changelog: v0.5.5...v0.5.6

Release: v0.5.5

Security

• Updated hono from 4.11.3 to 4.11.4 to fix JWT algorithm confusion vulnerability
  • See: GHSA-f67f-6cw9-8mq4
  • See: GHSA-3vhc-576x-3qv4

Release: v0.5.4

Security

• Update @modelcontextprotocol/sdk to 1.25.2 to fix ReDoS vulnerability in UriTemplate regex patterns