[PicoBerry]
PicoBerry MCP Server
Generate 3D models, images, and animations for game and 3D workflows from any
MCP client — Claude Code, Cursor, Claude Desktop, Cline — with no HTTP glue. A thin
wrapper over the PicoBerry /v1 API, so you get PicoBerry's
multi-engine pipeline directly inside your agent. Several 3D and image engines
sit behind one API; call list_models for the live set and each engine's cost.
Generated assets are drafts — useful for prototyping and iteration, and can be
reviewed or refined for your project.
📖 Full reference: API + MCP docs · PicoBerry API
There's no separate subscription for the MCP or the API. Generation spends
the same prepaid PicoBerry credits as the web app, per engine, at rates you can
read with list_models before you spend anything. (Using the API does require
a completed purchase — see Get an API key.)
Install
No install needed — run it with npx:
Cursor uses the same shape in ~/.cursor/mcp.json.
Get an API key
Sign in at https://picoberry.ai, open the API Keys
tab in your dashboard, and hit Create key. The key is shown once — copy it
immediately and treat it like a password.
API access needs a completed purchase: a subscription or a one-off credit
pack. A purchase entitles you permanently — you don't need a current
subscription. (An active paid subscription works too, of course.)
Environment variables
Tools
How generation works
Generation is asynchronous:
generate_3d_from_text({ prompt }) → returns an asset { id }.wait_for_asset({ asset_id: id }) → polls until taskStatus === 2 (succeeded).- Read the result URL from
files.model (GLB) or files.image (PNG).
taskStatus: 0 pending · 1 processing · 2 succeeded · 3 failed. Result
URLs are signed and short-lived — download promptly. Errors come back with an
actionable message (e.g. an unknown engine returns the list of valid names).
Example (in an agent)
"Make a low-poly treasure chest, retopo it to 3k tris, and give me a Unity FBX."
Use it alongside Blender MCP
Run this next to blender-mcp and the
agent can generate with PicoBerry, then import into Blender in one flow:
Develop
Release
Run Actions → Publish → Run workflow (or push a v* tag). It publishes to
npm and then to the official MCP registry, in that order — the registry
validates by fetching the package's npm metadata and matching its mcpName
against server.json's name, so npm has to land first. A guard step checks
every invariant (name/version agreement, namespace casing, version not already
on npm) before anything is published, because npm versions are immutable and a
failed half-publish burns the number.
Bump version in both package.json and server.json (version and
packages[0].version) — the guard fails the run if they disagree.
One-time setup — no secrets. Both publishes authenticate over the workflow's
GitHub OIDC token (id-token: write). There is nothing to store or rotate.
The only step is telling npm to trust this workflow. On npmjs.com go to
@picoberry/mcp-server → Settings → Trusted publishing → GitHub Actions and
enter:
The workflow filename must match exactly — it is part of what npm verifies.
The MCP registry needs no setup at all: mcp-publisher exchanges the Actions
OIDC token, and the registry grants io.github.<repository_owner>/* from the
token's repository_owner claim. That covers io.github.UModeler/picoberry-mcp
and avoids the interactive browser login (which additionally requires org Owner).
Trusted Publishing needs npm >= 11.5.1, so the workflow runs on Node 24
(npm 11.x). Node 22 still bundles npm 10.9 and would fail — the node-version
pin is load-bearing. A guard step fails the run early if the runner ever ships
an older npm.
The namespace is compared byte-exactly — io.github.UModeler/..., matching
the GitHub org's login. A lowercased io.github.umodeler/... is rejected 403.
After publishing, claim the Glama listing
— unclaimed servers get limited discoverability, and awesome-mcp-servers gates
its PRs on a Glama badge in CI.
License
MIT
