v0.2.1 — Imagen 4 deprecation

What's new

gemini_generate_image_imagen deprecated ahead of Imagen 4 sunset

Google discontinues all Imagen 4 endpoints on 2026-08-17 (imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001). After that date those model IDs return 404 Not Found.

gemini_generate_image_imagen now redirects every call to gemini-3.1-flash-image-preview (the same path used by gemini_generate_image) so no call can hit the 404:

• A call with no model= or with any imagen-4.0-* ID is served by gemini-3.1-flash-image-preview.
• An explicit non-Imagen model ID is still honored; GEMINI_DEFAULT_MODEL still applies.
• aspect_ratio is translated into a prompt instruction (the flash-image path has no aspect-ratio config field).
• count maps to candidate_count, matching gemini_generate_image behavior.
• Every response now carries "deprecated": true and a "deprecation" message.

For new code, use gemini_generate_image directly — it is the same flash-image path without the deprecated wrapper.

Migration guide

docs/migration-v0.2.1.md covers the redirect rules, capability translation, and the phase-2 removal timeline.

Other changes

• 74 unit tests (up from 69 in v0.2.0)
• aspect_ratio values are now validated against the allowed set before prompt interpolation
• GEMINI_DEFAULT_MODEL=imagen-4.0-* is redirected correctly (post-resolve guard)

v0.2.0 — I/O 26 refresh

What's new

Default model bumps (all existing tools)

Every tool's built-in default is now the newest model available on Google AI Studio (preview included). Users who relied on the v0.1.x defaults and want to pin will find a per-tool recipe in docs/migration-v0.2.md.

Tool	v0.1.x default	v0.2.0 default
gemini_generate	gemini-2.5-pro	gemini-3.1-pro-preview
gemini_generate_image	gemini-2.5-flash-image	gemini-3.1-flash-image-preview
gemini_generate_image_imagen	imagen-4.0-generate-001	imagen-4.0-ultra-generate-001
gemini_code_execute	gemini-2.5-pro	gemini-3.1-pro-preview-customtools
gemini_search_grounded	gemini-2.5-flash	gemini-3.5-flash
gemini_analyze_file	gemini-2.5-pro	gemini-3.1-pro-preview
gemini_chat	gemini-2.5-flash	gemini-3.5-flash
gemini_start_video	veo-3.0-generate-001	veo-3.1-generate-preview

The Veo flip also resolves the 2026-06-30 sunset of veo-3.0-generate-001.

Three new tools

gemini_generate_music — Text-to-music via Lyria 3. Returns a WAV file.

Tool: gemini_generate_music
  prompt: "upbeat lo-fi piano loop for studying"

gemini_tts — Text-to-speech with single-voice and multi-speaker support. Returns a WAV file.

Tool: gemini_tts
  text: "Hello from Gemini."
  voice: "Kore"

gemini_start_research + gemini_get_research_report — Deep Research polling pair. Start returns an operation_id; poll until done to get a Markdown report with citations. Runs in a background thread so the MCP server stays responsive during multi-minute synthesis.

Tool: gemini_start_research
  prompt: "What are the main tradeoffs between RAG and fine-tuning for LLM applications?"

Migration guide

docs/migration-v0.2.md lists every default flip and shows how to pin to the old model if you need stability.

Other changes

• 69 unit tests (up from 49)
• Updated infographic showing all 13 tools
• docs/migration-veo-3.1.md superseded by docs/migration-v0.2.md

v0.1.0 — gemini-mcp

Initial tagged release of gemini-mcp.

What it is

An MCP server exposing 10 Gemini tools (Gemini 2.5/3, Imagen 4, Veo 3) to Claude Code and any MCP-compatible client. See the README for install and usage.

Assets

• gemini-mcp-architecture.mp4 — a 7-minute architecture walkthrough of how Claude Code talks to Gemini through MCP.

Install

claude mcp add gemini -s user -e GEMINI_API_KEY=<your-key> -- uvx --from git+https://github.com/azmym/gemini-mcp@v0.1.0 gemini-mcp