Picsart MCP Server
OfficialProvides tools for generating images, video, and audio using Picsart's AI models across 141 models from 28 providers.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Picsart CLI & MCP — Docs
Public developer documentation for the Picsart gen-ai CLI, the Picsart MCP server, and drop-in Skills with an SDK catalog reference of 260 models from 32 providers. Built with VitePress.
Authoritative copy: the public repo, https://github.com/PicsArt/picsart-mcp-cli-docs — it accepts pull requests directly, so it can and does run ahead of the mirror below.
Mirror:
docs-site/inside the privateai-toolkitmonorepo. Catalog refreshes are usually authored here, but this folder goes stale the moment a PR lands publicly. Always reconcile before publishing — see Publishing.Themed site (live): https://picsart.github.io/picsart-mcp-cli-docs/
Live wiki (public, no Actions needed): https://github.com/PicsArt/picsart-mcp-cli-docs/wiki
Local development
cd docs-site
npm install
npm run dev # http://localhost:5173 (or --port 4600)
npm run build # static build → .vitepress/dist
npm run preview # preview the production buildRelated MCP server: AADDYY MCP Server
Structure
docs-site/
.vitepress/config.ts # nav, sidebar (unified), theme
.vitepress/theme/ # Picsart brand + ModelCatalog / ProviderGrid Vue components + data/
index.md # home (hero)
guide/ # getting started + CLI / MCP / Skills + concepts
reference/ # model reference: catalog, per-mode, per-provider
providers/ # one page per vendor (28)
public/llms.txt # generated AI-agent site map (llmstxt.org)
scripts/build-llms.mjs # generates public/llms.txt from the catalog data
scripts/check-counts.mjs # fails the build if prose model/provider counts drift
scripts/fix-counts.mjs # rewrites drifted counts (counterpart to check-counts)
scripts/draft-changelog.mjs # prints a draft changelog entry from the catalog diff
scripts/lib/counts.mjs # shared count patterns used by check/fix-counts
scripts/build-wiki.py # converts these docs → GitHub Wiki markdown
.github/workflows/refresh-catalog.yml # bump SDK → refresh → build → open PRData freshness
.vitepress/theme/data/{models,providers}.json are generated from the installed SDK catalog.
The development dependency pins @picsart/ai-sdk to 7.0.0.
Refreshing the catalog — one command. Bump the SDK pin in package.json, npm install, then:
export PICSART_API_KEY=<prod-key> # SDK 7.0.0+ fetches the catalog from models-hub; a .env is auto-read
npm run refresh:catalog # export → build-catalog-data → build-provider-pages → fix-counts
npm run draft:changelog -- --base HEAD # prints a changelog draft to paste into changelog.md
npm run build # verify (llms + check:counts + vitepress)refresh:catalog writes a transient .sdk-catalog.json (gitignored) and regenerates the Model
Catalog, Providers grid, provider pages, and wiki tables; fix:counts then rewrites any drifted
prose counts so check:counts passes. Still hand-edited after a refresh: the curated changelog
prose and the "SDK X.Y.Z catalog" baseline version strings.
Or let CI do it — fully hands-off. .github/workflows/refresh-catalog.yml runs the whole chain
daily (or on demand from the Actions tab): fetch the live catalog, regenerate, auto-fix counts,
write the changelog entry, open a PR, auto-merge it to main, and deploy to Pages. The
merge uses GITHUB_TOKEN (which can't trigger deploy-docs.yml by push), so the workflow dispatches
the deploy explicitly — no PAT needed. SDK 7.0.0+ is dynamic-only, so models added in models-hub
appear with no SDK release: leave the version input blank and the refresh re-fetches the live
catalog against the current pin (pass a version only to move the SDK itself). When nothing changed
it's a no-op — no PR, no deploy. The auto-written changelog is the raw catalog descriptions, not
editorial prose; tidy it in a follow-up if you care. Needs the PICSART_API_KEY secret and "Allow
GitHub Actions to create … pull requests" enabled.
To run the generator steps by hand instead: node scripts/export-sdk-catalog.mjs <output.json>
(needs PICSART_API_KEY), then pass that file to build-catalog-data.mjs and
build-provider-pages.mjs.
public/llms.txt (the llmstxt.org agent map) is generated from those
same JSON files, so it tracks the real model/provider counts. npm run build regenerates it
automatically; run it standalone with npm run llms. Like robots.txt/sitemap, it hardcodes
the production GitHub Pages subpath — override with DOCS_HOSTNAME / DOCS_BASE (e.g. on a
custom-domain move) and re-run.
The hand-written counts in prose (260 models, 78 image models, 32 providers, each
provider page's **Models:** N) are guarded by npm run check:counts, which recomputes the
truth from the same JSON and fails the build on any mismatch — so a stale count can't ship.
Run it standalone with npm run check:counts.
Publishing
The docs live in PicsArt/picsart-mcp-cli-docs (public), with two delivery surfaces: GitHub Pages and the GitHub Wiki. Both are published from the public repo, and both must be updated — Pages deploys itself, the Wiki does not.
Never copy docs-site/ over the public repo root. The public repo takes PRs
directly, so it routinely holds work this folder has never seen. As of 2026-08-25
that was an entire catalog refresh, 21 integration guides, 6 concept guides, the
community files, and a stricter check-counts.mjs. A blind mirror would have
deleted 32 tracked files, including .github/workflows/deploy-docs.yml — the
workflow that deploys the site.
Publishing a change: generate inside a clone of the public repo
Work in the public repo, not in docs-site/. That way the public content is the
base and your change is a diff on top of it, rather than a replacement of it.
git clone https://github.com/PicsArt/picsart-mcp-cli-docs.git /tmp/pubdocs
cd /tmp/pubdocs
# apply the change here — for a catalog refresh, one command does the chain:
npm ci
PICSART_API_KEY=<prod-key> npm run refresh:catalog # export → data → pages → fix-counts
npm run build # runs check:counts; VitePress fails on dead linksBefore committing, prove you are not deleting anything:
git status --porcelain | grep '^ D' && echo "STOP — this would delete files" || echo "safe: no deletions"
git add -A && git commit -m "docs: <what changed>" && git push origin mainThen sync the Wiki (Route B below) — Pages auto-deploys, the Wiki never does.
Finally, reconcile the mirror so the next refresh diffs against reality:
rsync -a --delete --exclude '.git/' --exclude 'node_modules/' \
--exclude '.vitepress/dist/' --exclude '.vitepress/cache/' --exclude '.wiki/' \
/tmp/pubdocs/ <ai-toolkit>/docs-site/Compute any model delta against the published
.vitepress/theme/data/models.json, never the local one. Diffing against a stale mirror silently miscounts what is new.
The
ai-toolkitmonorepo is private; never push its history to the public repo. Only thedocs-site/contents (at the public repo root) +.github/workflows/deploy-docs.ymlbelong there.
Route A — GitHub Pages (themed, interactive) ← live
.github/workflows/deploy-docs.yml auto-builds and deploys on every push to main.
It derives the base path from the repo name (DOCS_BASE=/<repo>/).
Actions are enabled (an org admin allowed them on 2026-06-18 — the PicsArt org disables
Actions by default) and Pages Source = GitHub Actions (build_type: workflow), so pushes
to main deploy automatically. To run a deploy by hand:
gh workflow run deploy-docs.yml --repo PicsArt/picsart-mcp-cli-docs --ref mainFallback if Actions is ever disabled again: build locally (
DOCS_BASE=/picsart-mcp-cli-docs/ npm run build) and push.vitepress/distto agh-pagesbranch with Pages set to "deploy from branch".
Route B — GitHub Wiki (plain, no Actions)
GitHub renders the wiki server-side, so it needs no Actions/Pages build. Plain Markdown, so the interactive catalog/provider grids become static tables.
First-page caveat: GitHub has no API to create a wiki's first page. Create one page via the web UI once (
/wiki→ "Create the first page" → Save). After that,.wiki.gitis pushable and the steps below keep it in sync.
# 1. clone the wiki (separate git repo)
git clone https://github.com/PicsArt/picsart-mcp-cli-docs.wiki.git .wiki
# 2. regenerate all wiki pages from these docs (guides + integrations + reference + providers + Home/_Sidebar/_Footer)
npm run wiki:build -- .wiki # = python3 scripts/build-wiki.py .wiki
# 3. publish
cd .wiki
git status --porcelain | grep '^ D' && echo "STOP — pages would be deleted" || echo "safe: no deletions"
git add -A && git commit -m "docs: sync wiki"
git push origin master # GitHub wikis use the 'master' branchscripts/build-wiki.py strips VitePress frontmatter, rewrites /guide and /reference
links to wiki page names, converts :::tip/warning containers to blockquotes, and replaces
the <ModelCatalog> / <ProviderGrid> components with static tables built from the data JSON.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Image, video, music and text generation across 100+ models through one endpoint.
Generate images, videos and audio across many AI models via one API, with quotes before you pay.
Generate images, video, music, voice and 3D through one API. 30 tools, 200+ models.
One API for 100+ AI video, image, music and speech models.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceProvides AI agents access to over 200 AI models and 10 service categories via a unified, metered API with transparent per-call pricing.9 npmMIT
- AlicenseCqualityCmaintenanceEnables AI agents to use over 100 AI tools for tasks like content generation, image creation, video generation, and problem solving via the AADDYY API.4914 npmMIT

Createya MCP & APIofficial
AlicenseNot gradedqualityDmaintenanceEnables AI agents and developers to generate images, videos, audio, and text using 100+ models via MCP or REST with a single API key.4MIT- AlicenseAqualityCmaintenanceEnables AI assistants to generate video, images, music, speech and sound effects, transcribe audio, edit and extend clips, and upscale media across 100+ models using a single API key. Also provides task tracking, model discovery, cost estimation and balance checks.14MIT