Scenario for Blender

Overview

Scenario for Blender is an open-source extension for Blender 5.0 and later, licensed GPL-3.0-or-later and marked experimental in its README. Users install it from the extension repository https://blender.scenario.com/repo/index.json (Edit > Preferences > Get Extensions > Repositories) or from a ZIP on the releases page; the handbook is blender.scenario.com. The Scenario API key and secret live in the extension's preferences: the agent never handles them. Blender's Allow Online Access must be on.

The extension runs a local MCP server, scenario-blender, at http://127.0.0.1:9876/mcp (JSON-RPC over HTTP POST with a bearer token, no SSE). When 9876 is busy it tries the next three ports, so the URL to use is the one shown in Scenario > Agents (MCP). Scene tools run on Blender's main thread. Every tool and parameter here comes from the extension's tool reference, docs/MCP.md; tools/list carries each tool's full guidance. Local names are verb-first (list_models, job_status) and never interchangeable with the hosted server's (models_list, model_run, jobs_wait). Training, collection rename or deletion, workflow authoring and usage stay on mcp.scenario.com: see the scenario skill, and connect both servers when a task needs both. Modeling, sculpting, rigging or shading through bpy is the scenario-blender-expert skill's work; exporting for an engine is scenario-blender-export's. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Setup

The token is generated the first time the server starts in a Blender session, held only in memory, and changes after every restart: a 401 means it is stale, so copy the setup again from the Agents (MCP) panel. It is a secret, separate from the API key: never paste it, or a copied snippet, into an issue, transcript or screenshot. The client must run on the same machine, since 127.0.0.1 is loopback, and must use that address: another host name returns 403 forbidden host.

Claude Code:

claude mcp add --transport http scenario-blender http://127.0.0.1:9876/mcp \
  --header "Authorization: Bearer <token>"

Codex:

export SCENARIO_BLENDER_TOKEN='<token>'
codex mcp add scenario-blender --url http://127.0.0.1:9876/mcp \
  --bearer-token-env-var SCENARIO_BLENDER_TOKEN

Cursor takes an mcpServers entry "scenario-blender": {"url": "http://127.0.0.1:9876/mcp", "headers": {"Authorization": "Bearer <token>"}}. Claude Desktop goes through the extension's stdio bridge: copy its snippet from the panel, which fills in the installed shim path, and check that its command points to a working Python 3 interpreter.

Headless, with the extension enabled:

export SCENARIO_BLENDER_TOKEN='<token>'
blender --background scene.blend --command scenario_blender --port 9876

--online-mode enables online access for that run. --token '<token>' overrides the variable but exposes the token in process arguments; with neither, the command generates one and prints it once. Headless, screenshot_viewport, render_still and every capture_reference source except MESH fail: they need the GUI.

Quick reference

In call order:

Step	Tool	Notes
Read the scene	scene_summary, datablocks_summary, object_detail(name)	Start here; scene_summary reports the selection, cameras, frame range and 3D cursor
Change the scene	select_objects(names), set_frame(frame), camera_path	camera_path takes a preset (such as orbit or push_in), a free-text description or waypoints, plus duration, focal, aim_at_subject; it animates the scene camera for Render Video and generates nothing
Look	screenshot_viewport, render_still(source, width, height)	GUI only; PNG content for the agent, not an upload
Reference	capture_reference(source), then reference_upload_status(context_id, reference_id)	source: VIEWPORT, CAMERA, VIEWPORT_CLIP, CAMERA_CLIP or MESH (selected meshes as one GLB). Poll until state is imported, then use its asset_id
Pick a model	list_models(lane, query), model_schema(model_id)	Lanes: image, video, 3d, material, audio, render_image, render_video, edit3d, panorama. Ids come from list_models each session, never constants
Quote	estimate_cost(model_id, parameters, lane)	Free dry run; returns quote_id and cu_cost_exact (a decimal string). lane defaults to image
Generate	generate(lane, model_id, parameters, quote_id, approved_cost)	Spends credits. Same lane, model and parameters as the quote; approved_cost is the exact string the user approved
Wait	wait_for_job(job_id, timeout), job_status(job_id)	timeout defaults to 45 seconds, inside the 60-second limit many clients apply; go up to 170 only when your client waits longer. Call again while delivery_active is true
Place	list_local_jobs, prepare_result_application, then apply_result_application(context_id, application_id)	Only Image results import by themselves; everything else waits for an approved destination
Python	blender_api_help(path), then execute_python(code)	Off until the user enables Allow connected agents to run Python

upload_reference(path, kind) sends only a file the user chose in Blender this session or one the extension saved; any other path is refused, so capture the scene instead or ask the user to pick the file. Render Image and Render Video take no raw parameters: render_form with action: "configure", then action: "prepare" captures and uploads the scene, then estimate_cost and generate with parameters omitted. The model copies the capture's motion, limbs included: for characters that must walk or run naturally, block them as capsules (scenario-blender-blockout). A panorama goes into the World through prepare_result_application with purpose: "world".

Worked example: viewport to a mesh at the 3D cursor

"Turn what I see in the viewport into a 3D prop next to it."

1. scene_summary: note cursor and confirm a GUI session, since the capture needs one.
2. capture_reference with {"source": "VIEWPORT"} returns context_id and reference_id, not an asset id. Call reference_upload_status with both until state is imported, and keep its asset_id. A scene change before the upload finishes stops it.
3. list_models with {"lane": "3d"}; pick an image-to-3D entry from its description. model_schema with its model_id names the image input; file parameters take Scenario asset ids.
4. estimate_cost with {"lane": "3d", "model_id": "<id>", "parameters": {"<image input>": "<asset_id>"}} returns quote_id and cu_cost_exact. Tell the user the cost in CU and stop until they say yes. Unattended, with nobody to answer, report the quote and stop here: neither this gate nor the one in step 8 is ever passed without the user.
5. generate with the same lane, model_id and parameters, plus quote_id and approved_cost set to cu_cost_exact verbatim. It returns local_id.
6. wait_for_job with {"job_id": "<local_id>"} (45 seconds by default), again while delivery_active is true. The 3D lane stops at saved results: nothing is in the scene yet.
7. job_status lists import_units; take the key of the primary unit. list_local_jobs gives context_id and this job's request_id and revision.
8. prepare_result_application with {"context_id": "...", "request_id": "...", "expected_revision": <revision>, "asset_id": "<key>"} captures the current 3D cursor as the destination and returns application_id. Show the user the returned scene, cursor, format and files, and wait for approval.
9. apply_result_application with context_id and application_id imports one new model group at that cursor. scene_summary confirms it.

Common mistakes

• generate without a fresh estimate_cost and the user's yes. A changed lane, model, input, scene or credential needs a new quote, a quote is used once, and an estimate never authorizes spending by itself.
• Repeating a timed-out generate. Inspect job_status or list_local_jobs instead; an unconfirmed submission blocks an identical one until the user marks it reviewed in Blender.
• Passing a file path, or capture_reference's reference_id, where a file parameter takes an asset id.
• Expecting import_result to place a result: it only explains the approval flow above.
• Captures in --background: only capture_reference MESH works there.
• Assuming execute_python is on. It fails with PermissionError naming Scenario preferences (MCP > Allow connected agents to run Python), off by default: ask the user, and prefer the specific tools.
• Calling the hosted server's tools (model_run, jobs_wait, model_schema_get) on this server.
• Treating "catalog is still loading" or "model description is still loading" as a failure: nothing was quoted, submitted or changed, so call again. Missing credentials in the extension's preferences keep the catalog from loading.
• Assuming the file is saved: no tool saves the .blend, so ask the user to save it.