Key takeaways
- Feed agents deterministic contracts: give coding agents report.json, scene.json / collision.json, and checksums.json so camera framing, grounding, and doorway clearance use measured numbers.
- Prototype offline for $0 first: use SceneCrate’s nine CC0 downloadable zips to wire up Three.js loading, LOD switching, and collision tests before spending any generation credits.
- Scoped Streamable HTTP MCP: SceneCrate’s /mcp endpoint exposes four focused tools (get_service_status, upload_image, create_asset, get_asset) authenticated with a personal 30-day paid-account Bearer key.
- Strict safety boundaries: explicit confirm_generation: true, mandatory idempotency_key reuse on retries, returned credits on failed technical checks, and immediate key revocation.
1. Give coding agents measured JSON contracts, not binary guesswork
Language models cannot eyeball a binary .glb buffer to tell whether a cabin door is 1.64 m wide, whether a prop’s origin floats 0.4 m above the floor, or whether a terrain mesh uses Z-up or Y-up coordinates. When left to guess, agents hardcode arbitrary scale multipliers (0.01 or 10) and wrap enterable buildings in single bounding boxes that block doorways.
Every SceneCrate scene and static-prop delivery pairs its .glb files with machine-readable JSON files: report.json records exact triangle, vertex, and material counts alongside Y-up metre bounding boxes; scene.json or collision/collision.json defines explicit walk boxes or 49 × 49 heightfields; and checksums.json lists SHA-256 digests. Pointing your coding agent at those JSON contracts lets it place spawns, cameras, and physics colliders accurately on the first pass.
2. Prototype locally with CC0 packs before calling paid tools
Do not debug your Three.js asset loader by burning paid generation calls. Download one of the nine free CC0 packs (such as Shipping Crate for single-box props, Cedarwatch Cabin for doorway walk boxes, or Highland Trail for heightfield sampling) and ask your coding agent to integrate standard.glb, lightweight.glb, and scene.json locally first.
Have your agent add automated unit checks verifying GLB header integrity, bounding-box height in metres, and WebGL resource disposal on scene teardown. Once the local loop is green, replacing a placeholder crate with a custom static prop becomes a predictable data swap.
3. How SceneCrate’s scoped MCP connection works
When you have an active paid credit pack and the service is open, you can issue a personal Bearer key on the Connect your agent (/agents) page and add SceneCrate’s stateless Streamable HTTP endpoint (/mcp) to your MCP-compatible coding agent.
The endpoint exposes four explicit tools: get_service_status (read-only check of live availability, text/image mode, and credit balance), upload_image (uploads a base64 PNG or JPEG up to 4 MiB with an idempotency_key, returning a file_key with zero arbitrary URL fetching), create_asset (creates one static prop from prompt or image_key with height_meters between 0.01 and 10, requiring confirm_generation: true and idempotency_key), and get_asset (reads task status, /studio preview link, and 5-minute authenticated download URLs).
{
"mcpServers": {
"scenecrate": {
"url": "https://scenecrate.com/mcp",
"headers": {
"Authorization": "Bearer YOUR_AGENT_KEY"
}
}
}
}4. Credit, privacy, and operational guardrails
Automated agent loops can easily double-spend if an HTTP timeout triggers a blind retry. SceneCrate’s upload_image and create_asset tools require an 8 to 100 character idempotency_key: if a network read times out, instruct your agent to resume with the exact same idempotency_key or poll get_asset rather than generating a new key. Each accepted creation reserves one credit, and any task that fails a technical geometry or budget check automatically returns that credit.
Agent keys belong strictly to individual paid accounts: unpaid accounts, admin-only test grants, and refunded packs cannot issue or retain active MCP keys. Keys are displayed once at creation, expire automatically after 30 days, and can be revoked in one click when a sprint ends. Connecting via MCP never bypasses paused service flags, reference-image privacy consent updates, or static convex-prop limits.
- Store YOUR_AGENT_KEY in local secret settings or environment variables; never commit it to git.
- Call get_service_status first to confirm creation_available and remaining credit balance.
- Reuse the exact same idempotency_key when retrying an interrupted upload or create_asset call.
- Revoke the personal agent key from /agents or /account as soon as your workflow finishes.
Frequently asked questions
What MCP tools does SceneCrate provide to coding agents?
SceneCrate exposes four tools over Streamable HTTP (/mcp): get_service_status (read-only status and credits), upload_image (PNG/JPEG up to 4 MiB), create_asset (explicit static prop creation with metre height), and get_asset (status, studio preview link, and 5-minute authenticated download URL).
How does SceneCrate prevent coding agents from double-charging on retries?
Both upload_image and create_asset require an explicit idempotency_key (and confirm_generation: true on creation). Re-sending the same idempotency_key resumes the original intent instead of reserving a second credit.
Can I use the MCP endpoint on a free or unpaid account?
No. Personal MCP keys are only issued to accounts holding a valid paid credit pack, expire after 30 days, and are deactivated if access is refunded or revoked. You can prototype agent workflows for free offline using the nine CC0 scene zips.
Does MCP access allow custom buildings, terrain, or rigged characters?
No. The MCP endpoint uses the exact same backend availability checks, privacy consent rules, and closed convex static-prop delivery limits (<= 15,000 / <= 5,000 triangles) as the web studio.