PicoBerry MCP Server is described as an AI 3D workspace for gaming, VR, and beyond. Its core goal is to generate, remesh, texture, and animate 3D assets in any MCP client, achieving "multiple engines, one API."
Purpose This tool is aimed at game and 3D workflows, and can directly generate 3D models, images, and animations in MCP clients (such as Claude Code, Cursor, Claude Desktop, Cline) without writing HTTP glue code. It is a lightweight wrapper around the PicoBerry /v1 API, making PicoBerry's multi-engine pipeline directly usable inside an agent.
Main capabilities It supports generating 3D models from text prompts (such as the example "generate a low-poly treasure chest, retopologize to 3k triangles, and give me a Unity FBX"), and also supports starting from a reference image. After generation, it can perform remeshing, retexturing, automatic rigging, and animation creation, and export to GLB, FBX, OBJ, and other formats. Multiple 3D and image engines sit behind the same API, and list_models can be called to get the live engine set and the cost of each engine. Generated assets are positioned as drafts, suitable for prototyping and iteration, and can be reviewed or refined for a project.
How to use No installation is required; it can be run directly with npx. In Cursor, write the same configuration structure into ~/.cursor/mcp.json. First, https://picoberry.ai register and log in, open the API Keys tab in the dashboard, and click "Create key." The key is shown only once, so copy it immediately and keep it safe like a password. API access requires completing a purchase (subscription or one-time credit pack); after purchase, authorization is permanent and there is no need to keep the current subscription active. Using the API requires completing a purchase, see "Get an API key" for details. There is no separate subscription for MCP or the API; generation consumes prepaid PicoBerry credits the same as the web app, billed by engine, and rates can be viewed before spending via list_models.
Generation flow (asynchronous)
- Call generate_3d_from_text({ prompt }) → returns an asset { id }.
- Call wait_for_asset({ asset_id: id }) → poll until taskStatus === 2 (success).
- Read the result URL from files.model (GLB) or files.image (PNG). Where taskStatus: 0 pending, 1 processing, 2 success, 3 failed. Result URLs are signed and short-lived, so download them promptly. Errors are returned with actionable information (for example, an unknown engine returns a list of valid names).
Working with Blender MCP It can run alongside blender-mcp, allowing an agent to first generate with PicoBerry and then import into Blender in one flow.
Release-related (for code contributors) Run Actions → Publish → Run workflow (or push a v* tag), and it will first publish to npm, then to the official MCP registry. The order cannot be reversed, because the registry validates by scraping the package's npm metadata and matching mcpName with the name in server.json, so npm must complete first. Before publishing there is a guard step that checks all invariants (name/version consistency, namespace casing, version not already present on npm), because npm versions are immutable, and a failed half-publish will consume that version number. The version must be bumped in both package.json and server.json (version and packages[0].version). If bothAn inconsistency guard will cause the run to fail. One-time setup requires no keys: both publishes authenticate via the workflow's GitHub OIDC token (id-token: write), with nothing to store or rotate. You only need to let npm trust the workflow on npmjs.com: go to @picoberry/mcp-server → Settings → Trusted publishing → GitHub Actions and enter the corresponding information; the workflow filename must match exactly, because it is part of npm's verification. The MCP registry requires no setup at all: mcp-publisher exchanges the Actions OIDC token, and the registry grants io.github.<repository_owner>/* based on the token's repository_owner claim, covering io.github.UModeler/picoberry-mcp and avoiding the need for an interactive browser login by an organization Owner. Trusted Publishing requires npm >= 11.5.1, so the workflow runs on Node 24 (npm 11.x); Node 22, which still bundles npm 10.9, will fail, making the node-version pin critical. Namespaces are compared byte-for-byte: io.github.UModeler/... matches the GitHub organization's login, while lowercase io.github.umodeler/... will be rejected with a 403. After publishing, claim the Glama listing—unclaimed servers have limited discoverability, and awesome-mcp-servers uses the Glama badge in CI as the threshold for its PRs.
Limitations The generated assets are drafts, suitable for prototyping and iteration. The API requires completing a purchase. Result URLs are signed and short-lived, so they must be downloaded promptly. Specific values for details such as environment variables are not provided in the materials.
