mcp server
ideogram-ai-toolkit
Claude Code skill for Ideogram AI image generation — structured JSON caption prompting, color palette control, and MCP-based style extraction from reference images.
Set up
Runs on your computer. Pick your agent and operating system.
This project does not publish an install method we can turn into a config. Follow the README instead.
README
synced Oct 26, 2026ideogram-ai-toolkit
A set of Claude Code skills for Ideogram AI image generation — general prompting technique (loose vs. structured, color-palette control, reference-image style extraction, remix vs. edit), plus purpose-built skills covering brand systems, characters, marketing creative, icon sets, patterns/textures, bulk generation, custom model training, collections, and post-processing (background removal, upscaling). Built by Devkind.
Ideogram 4 was trained on structured JSON captions, not plain text — this toolkit documents that schema (adapted from the official ideogram-oss/ideogram4 docs, Apache-2.0) and packages practical prompting technique around it, scoped to what the connected Ideogram MCP actually supports.
What's in here
skills/ideogram-prompt/ # general Ideogram prompting technique
├── SKILL.md
├── references/ # JSON caption schema, style-extraction workflow
└── examples/ # pointer to the root examples/ set (see below)
skills/logo-prompting/ # logo/brand-mark prompt writing, Ideogram-first
├── SKILL.md
├── references/ # anti-slop discipline, brand visual-layer reading, intake checklist
└── examples/ # accumulated real logo-prompt directions by style family
skills/moodboard-generator/ # pre-logo 3x3 brand-exploration board (loose adjectives, unlocked palette)
├── SKILL.md
├── references/ # panel anatomy, composition-spec JSON schema, anti-slop discipline
└── examples/ # a worked prompt + full JSON breakdown
skills/concept-grid-generator/ # cheap comparison grid of distinct motif concepts, locked palette + technique, before a full board
├── SKILL.md
├── references/ # anti-slop discipline (reused from brand-identity-sheet)
└── examples/ # a worked prompt + handoff note
skills/brand-identity-sheet/ # whole brand system in one generated image (wordmark, icons, buttons)
├── SKILL.md
├── references/ # panel anatomy, composition-spec JSON schema, anti-slop discipline
├── examples/ # a worked prompt + full JSON breakdown
└── evals/ # eval scenarios for this skill
skills/character-model-sheet/ # multi-panel character turnaround sheet (mascot, game/animation character)
├── SKILL.md
├── references/ # panel anatomy, composition-spec JSON schema, anti-slop discipline
├── examples/ # a worked prompt + full JSON breakdown
└── evals/ # eval scenarios for this skill
skills/marketing-ui-mockup-hero-generator/ # landing-page hero shots, device mockups, dashboard screenshots
├── SKILL.md
├── references/
├── examples/ # worked hero/device-mockup/dashboard examples
└── evals/
skills/social-marketing-asset-generator/ # finished social/ad creative with copy baked into the pixels
├── SKILL.md
├── references/ # composition-spec format, anti-slop discipline, platform dimension map, typography-baking discipline
├── examples/ # populated as worked jobs are run — see examples/STATUS.md
└── evals/
skills/photographic-icon-set-generator/ # consistent 3D/claymation/textured icon packs sharing a locked style recipe
├── SKILL.md
├── references/ # style-lock recipe, anti-slop discipline, set-consistency workflow
├── examples/ # a worked claymation icon set
└── evals/
skills/seamless-pattern-texture-generator/ # seamless/tileable patterns and material textures
├── SKILL.md
├── references/ # tile-prompt recipe, tiling verification, anti-slop discipline, composition-spec format
├── examples/ # populated once a real job is run — see examples/README.md
└── evals/
skills/bulk-image-generation-workflow/ # one locked caption → N on-brief variations, batch-submitted and culled
├── SKILL.md
├── references/ # variation strategy, review/culling guide
├── examples/ # a worked 10-prompt sticker-pack batch, shortlisted to 6
└── evals/
skills/custom-model-training/ # train a custom Ideogram model on a folder of reference images
├── SKILL.md
├── references/
├── examples/ # a worked training set + proof generation
└── evals/
skills/world-builder/ # expand a trained custom character model into a whole brand world
├── SKILL.md
├── references/ # palette lock, anti-slop discipline, character-batch discipline, batch tracking, contamination check
├── examples/ # populated once a real world-building pipeline is run — see examples/README.md
└── evals/
skills/collections-management/ # create/browse/rename/delete Ideogram collections, file images into them
├── SKILL.md
├── references/
├── examples/ # a worked logo collection
└── evals/
skills/remove-background-workflow/ # strip an image's background to a transparent PNG
├── SKILL.md
├── references/ # background-removal patterns
├── examples/ # a worked before/after
└── evals/
skills/upscale-image-workflow/ # raise the resolution of one specific Ideogram image
├── SKILL.md
├── references/ # upscale settings
├── examples/ # a worked upscale
└── evals/
skills/reframe-image-workflow/ # reframe an existing Ideogram image to a new aspect ratio
├── SKILL.md
├── references/ # source and ratio resolution
└── examples/ # a worked 9:16 logo reframe
examples/ # generated showcase assets + the exact prompts used
See examples/ for the full write-up of the root showcase set — each image below links to its exact prompt.
Install
This is a set of Claude Code skills. Requires the Ideogram MCP to be connected (claude mcp add ideogram --transport http https://mcp.ideogram.ai/mcp).
Each skill installs independently — swap ideogram-prompt below for any of the 17 skills under skills/ as needed.
As a Claude Code plugin (installs all 17 skills at once; needs Claude Code 2.1.275 or later):
/plugin install ideogram-ai-toolkit --marketplace devkindhq/ideogram-ai-toolkit
Via the skills CLI (reads straight from this GitHub repo, no unreviewed single-URL installer):
npx skills add https://github.com/devkindhq/ideogram-ai-toolkit/tree/main/skills/ideogram-prompt
# or, from a project root, target it explicitly:
npx skills add devkindhq/ideogram-ai-toolkit -s ideogram-prompt
Add -g to install globally (available in every session) instead of just the current project.
Project-scoped, manual (checked into a repo, shared with your team):
git clone https://github.com/devkindhq/ideogram-ai-toolkit.git /tmp/ideogram-ai-toolkit
cp -r /tmp/ideogram-ai-toolkit/skills/ideogram-prompt <your-project>/.claude/skills/ideogram-prompt
Personal, manual (available in every session):
git clone https://github.com/devkindhq/ideogram-ai-toolkit.git ~/ideogram-ai-toolkit
ln -s ~/ideogram-ai-toolkit/skills/ideogram-prompt ~/.claude/skills/ideogram-prompt
We deliberately don't ship a one-line curl ... | skill.md installer — pulling an unreviewed instruction file straight into ~/.claude/skills/ from a single URL is a real prompt-injection / supply-chain risk for an AI agent, and we'd rather you read the skill before trusting it. Clone it, read SKILL.md, then symlink or copy it in.
Network access and data
This plugin ships static skill files (markdown, JSON examples and sample images). It has no code, hooks, scripts or MCP server of its own, and it collects nothing.
The skills drive your own Ideogram MCP connection. Prompts and any images you upload go to Ideogram under your own account and subject to Ideogram's terms. You set up and hold the Ideogram MCP connection and its credentials yourself; this plugin does not read, store or request them.
Some skills tell Claude to run the single-use curl upload command that Ideogram's upload_image tool returns, or to download a generated image from an ideogram.ai URL, so a local file can reach Ideogram or a result can be saved. The skills also reference local jq and sips for filtering tool output and converting WebP to PNG.
Why this exists
Some public prompting guides for Ideogram reference a one-line curl installer for a Claude Code skill hosted at a single URL outside GitHub. That URL wasn't independently verifiable at the time this toolkit was built (returned HTTP 403, no cached or archived copy found), so this repo exists as an open, reviewable alternative: same underlying technique (documented against the real, open-source ideogram-oss/ideogram4 schema and the actual Ideogram MCP tool surface), but as source you can read before you run it.
License
Apache License 2.0 — see LICENSE and NOTICE. The JSON caption schema reference is adapted from ideogram-oss/ideogram4 (Apache-2.0).