io.github.loopcut-aiPublished by GitHub account loopcut-aiv0.3.0Updated Oct 30, 2026
Loopcut's cloud tools for Blender: GPU render, assets, images, 3D models, motion and audio.
Loopcut gives an assistant paid cloud tools for Blender work: GPU renders, free asset search, images, 3D models, rigging, motion and audio, billed in Loopcut…
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Follow the launch instructions in the repository.
The connector for Loopcut, an open-source AI agent inside Blender: cloud render,
assets, images, 3D models, motion and audio for Claude Code, Codex, Cursor, Claude Desktop and the shell, and
as a hosted connector for claude.ai and ChatGPT. Render .blend files on cloud RTX 4090s, search free asset
libraries, make and edit images, turn images into 3D models, remesh meshes, rig characters, make sound effects and voice
lines, and turn a video or a sentence into character animation on your own rig. Results come back as files the agent imports with Blender's Python. Everything is paid from your
Loopcut account's credits. You don't need the Loopcut app; any Blender install packs the scene.
It's one MCP server plus a CLI with the same verbs. The tools come from Loopcut's catalog, so a cloud tool that opens on Loopcut shows up without a new release.
Results are files the next tool takes: a generated image as image_to_3d's image, a model as remesh's
model, a rendered frame as image_edit's image. With a budget from you, the agent passes max_credits to a
paid tool, which starts a quote within it at once instead of returning it for job_start.
A character from a photo, rigged and animated: image_edit the photo into a T-pose with separated fingers,
image_to_3d from that with topology: "quads", import it and save the .blend, auto_rig it (20 credits),
then motion_from_text and motion_apply. auto_rig takes humanoids with clear arms, legs and fingers, up to
100,000 triangles; a rig that can't be bound is refunded automatically.
Two motion tools, with a free local tool to use their results:
motion.bvh is a rig-neutral SOMA 77-joint skeleton (standing T-pose rest, centimetres, Y up, 30 fps).
The connector measures the clip's length itself (MP4 and MOV from the file header, WebM with ffprobe
when it's installed) because the quote holds the seconds captured. target_armature and in_place on a
motion tool are kept on your machine for motion_apply and never sent to Loopcut. job_status returns
the result with three ways to use it: motion_apply (or loopcut motion-apply) for a saved .blend,
a short Python snippet for a Blender that's already open (for example through another Blender MCP),
and Blender's plain BVH import. The retargeting is the Loopcut add-on's own, so results match the app.
With no-generation on, motion_from_video stays available (it records the user's own performance) and
motion_from_text is hidden.
Price. 1 credit = $0.01. Every paid tool quotes first, for free, and the quote is the most a job can cost. Rendering costs 5 credits per GPU-minute on an RTX 4090 with a one-minute minimum; the other prices are on loopcut.org/pricing. Asset search, quotes, packing and the render QA report are free. Failed jobs are refunded. New accounts start with 100 credits.
You need uv (it provides Python 3.10+), and Blender if you render or apply a
motion. To get uv: brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh on macOS and Linux,
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex" on Windows.
claude.ai, ChatGPT and Claude Desktop, nothing to install: add https://loopcut.org/mcp as a connector and
sign in with Loopcut. In Claude: Customize → Connectors → Add custom connector. In ChatGPT: chatgpt.com/plugins →
+ → Add custom MCP server (custom servers with paid tools need a Business, Enterprise or Edu workspace
until Loopcut is in ChatGPT's plugin directory). The hosted connector has asset search, images, 3D models,
remeshing, motion from text, sound, voice and rendering from a script, and returns download links; one tool's
result feeds the next by its job:<job_id>/<name> ref. Rendering your own .blend and motion from a video
need the plugin below, because they read files on your computer.
Claude Code (plugin, which includes a skill that teaches the agent the tools and the habits that save credits)
On Claude Code before 2.1.275, run /plugin marketplace add loopcut-ai/loopcut and then
/plugin install loopcut@loopcut. Or add only the MCP server:
Codex (plugin, with the same skill)
Or add only the MCP server:
Every tool call returns within 50 s, under Codex's default 60 s tool timeout; long work carries on between
calls. Both plugins start the server with uvx loopcut==<release> mcp, pinned to the release they ship with; the
package has no dependencies beyond Python.
Cursor, VS Code, Claude Desktop and other MCP clients
Apps opened from the Dock or Start menu may not see your shell's PATH. If the server doesn't start, use
the full path from which uvx as command.
The first call opens a browser to sign in, and the key is kept where only you can read it. Without a
browser (over SSH, say), the call returns the link for you to open. LOOPCUT_API_KEY skips all of this.
Install the command with uv tool install loopcut, or run any command without installing as uvx loopcut ….
run and render quote, then start only with --yes, under --max-credits N, or after you answer the
prompt. run of a motion tool with --blend also puts the motion on that scene's rig; motion-apply
saves over --blend (Blender keeps the previous file as .blend1) unless you give --out. Results print as JSON on stdout; progress goes to stderr. In CI, create a key on
loopcut.org/account and set LOOPCUT_API_KEY.
Private mode and no-generation. These account settings apply here too. Private mode (on by default on
paid plans) refuses tools whose provider may train on what it receives; today that is sound effects.
No-generation hides and refuses the generative tools. Change them in the Loopcut app or with
loopcut settings.
-b --factory-startup --disable-autoexec) opens the saved
.blend and runs the Loopcut add-on's own collector (addon/bundles.py, vendored). It copies every
texture, library, cache and sequencer file into a bundle with relative paths. That Blender gets a
private config folder and none of your environment's keys. The cloud never runs Python from your file,
and the renderer has no internet access.loopcut_renders/<job_id>/
next to the .blend for a render, and loopcut_outputs/<job_id>/ in the working folder for everything
else (or out_dir). The working folder is the client's workspace root when it shares one; Codex starts
plugin servers in the plugin's own folder, so there results go under your home folder instead. A
render's QA report flags black, white, flat and transparent frames, fireflies, duplicate frames and
exposure jumps.motion_apply starts the same kind of background Blender on your saved .blend
and runs the add-on's retargeter (addon/motion_retarget.py, vendored) through
loopcut/motion_blender.py: one whole-body rest alignment, limb-only direction alignment for A-pose
rigs, the pelvis path scaled by leg length from the character's own standing height, and keys written
parent first. A previous action on the rig is kept in a muted NLA track. Nothing leaves your machine and it
costs nothing.auto_rig runs the add-on's rigging code (addon/autorig.py and rig_*.py, vendored) in a
background Blender through loopcut/rig_blender.py, once on each side of the cloud job: first it checks the
named meshes and writes the upload (the meshes as a GLB and a rig task with their fingerprints) to the state
folder, without changing the .blend; after the job it checks the character is unchanged, verifies the
returned skeleton and weights against what was sent, binds them to the original meshes and saves.If the catalog can't be read at startup, the server uses the last catalog it read from that service, then
the snapshot in src/loopcut/catalog.json.
What it runs and connects to. uv, which provides Python (and downloads it if needed); your Blender,
headless, only to pack a scene for render_quote, to apply a motion with motion_apply and to check and bind a rig for auto_rig; ffprobe, if
installed, only to measure a WebM clip; and your browser, only to open the Loopcut sign-in
page. It connects to loopcut.org (the catalog, sign-in and job API) and to the presigned storage URLs
that API hands out for uploads and downloads. It sends no analytics.
Loopcut's privacy policy is at loopcut.org/privacy. For this connector:
.blend and every texture, library and cache it uses. For other tools, the
inputs you or the agent give (prompts, settings) and the files named as inputs (images, meshes, videos).
It reads nothing else from your machine: no other files, no environment keys besides LOOPCUT_API_KEY,
no open Blender session. motion_apply opens and saves the .blend you name on your machine only, and
target_armature and in_place stay local too. For auto_rig it sends the named meshes (as a GLB) and their
fingerprints, not the .blend, and binds the result on your machine.~/.config/loopcut-connector, readable only by you. loopcut logout removes the key.The end-to-end tests run against Loopcut's contract mock, which lives in the service repository; they're
skipped here. addon/VENDORED.md explains how the add-on copies are kept in sync.
Released under the GPL-3.0-or-later. Loopcut is an independent project; Blender is a trademark of the Blender Foundation.
Source: README.md at commit 2d7bb4e
Sign in to Loopcut
Sign in to the user's Loopcut account, or switch accounts. Any tool called while signed out does this by itself. Opens the sign-in page in the browser and waits; if it returns a verification_url instead, show it to the user and call again after they approve.
wait_secondsLoopcut account
The signed-in Loopcut account: plan, credit balance, Private mode and no-generation settings, and which tools it can use.
Takes no parameters.
Cloud rendering
Price a cloud render of a saved .blend on an RTX 4090. Packs the scene and every texture, library and cache it uses with a local headless Blender and uploads what the cloud lacks. While the result says status "preparing", call again with the same arguments. Frames default to frame 1 only. Check warnings for missing files (they render pink). Returns a quote and spends nothing; call job_start with the quote_id once the user agrees, or pass max_credits (the user's budget) to start at once.
blendrequiredcameraengineblenderdenoisesamplesframe_endturntableframe_stepframe_startmax_creditswait_seconds+2 moreLoopcut cloud asset library
Search free CC0 libraries (Poly Haven, ambientCG, Physically Based) for models, materials, HDRIs and textures, with licence, size, preview and download details. Try it before generating a model or texture. Free: runs at once and returns the results.
kindlimitqueryrequiredlicensesourcesmax_polycountmin_resolutionImage to 3D
Make a textured 3D model (GLB) from a prompt, a local reference image, or both. For new props and concept meshes, not exact CAD. Takes 1-5 minutes. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
imagepromptqualitytexturedtopologymax_creditsconfirm_tokenRetopology
Cloud remesh: rebuild a mesh with clean quads or triangles at a target face count, with its textures baked onto the new mesh. Export the object to a .glb first. Quads come back as FBX, triangles as GLB. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
modelrequiredtopologymax_creditstarget_facesconfirm_tokenAuto-rig
Rig a humanoid character in a saved .blend: a skeleton with Mixamo bone names and skin weights, bound to the original meshes, ready for motion_apply. 20 credits. A local Blender checks it first, free: clear arms, legs and fingers, at most 100,000 triangles. The finished job saves the rig into the .blend (old one kept as .blend1). Use it instead of rigging by hand. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
blendrequiredblendertargetsmax_creditswait_secondsconfirm_tokenaccept_experimentalSound effects
Make a sound effect (WAV, 0.5-30 s) from a description: impacts, ambience, foley. Not speech or music. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
looppromptrequiredmax_creditsconfirm_tokenduration_secondsVoice lines
Speak a line of text as WAV audio with a chosen voice and direction, for characters or narration. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
textrequiredvoicelanguagedirectionmax_creditsconfirm_tokenImage generation
Make a PNG image from a prompt, optionally guided by a local reference image: concept art, backplates, references. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
sizepromptrequiredqualityreferencebackgroundmax_creditsconfirm_tokenImage editing
Change an existing image (a local PNG, JPEG or WebP) from a text instruction, with an optional PNG mask, e.g. to restyle a render. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
masksizeimagerequiredpromptrequiredqualitymax_creditsconfirm_tokenMotion from video
Capture a person's full-body motion, hands included, from a local MP4, MOV or WebM (one person, whole body in view, mostly still camera; at most 60 s captured, chosen with start_s and end_s). 3 credits per second captured, 2-second minimum. Returns motion.bvh (SOMA 77-joint skeleton) and motion.json; motion_apply then puts it on the user's rig in a .blend. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
end_svideorequiredstart_sin_placemax_creditsconfirm_tokenstatic_cameravideo_secondstarget_armatureMotion from text
Make 2-10 s of humanoid body animation from a short description of one simple physical action (NVIDIA Kimodo, built with Meta Llama 3). 2 credits per second of motion. Good at grounded motion (walks, dances, kicks, gestures, fighting in place); flips, spins in the air and big jumps come out small or missing. Say what the body does, not choreography: each sentence becomes its own segment, so one short sentence per action and no sentence of style words. Fingers stay still; not for faces, props or finger work. Returns motion.bvh (SOMA 77-joint skeleton) and motion.json; motion_apply then puts it on the user's rig in a .blend. Report what the motion does, not what was asked; if it misses, run again with a simpler prompt (charged again) or say so, never keyframe the root or hips over it to fake missing motion. Returns a quote and spends nothing; call job_start once the user agrees, or pass max_credits (the user's budget) to start at once.
seedpromptrequiredin_placeduration_smax_creditsconfirm_tokentarget_armaturePut a motion on a rig
Retarget a motion_from_video or motion_from_text result (motion.bvh) onto the humanoid rig in a saved .blend (Mixamo, Rigify, Auto-Rig Pro, Unreal mannequin or any humanoid by bone names) with a local headless Blender, then save the file; the previous version stays as .blend1. Free, nothing leaves the machine. target_armature and in_place default to what the motion tool was called with; without an armature it uses the active, the selected or the only humanoid armature, and with none it adds SOMA's skeleton with a simple body. If the .blend is open in Blender, run the job result's apply.bpy in that session instead. While the result says status "applying", call again with the same arguments.
bvhrequiredoutblendrequiredblenderin_placeframe_startwait_secondstarget_armatureStart a quoted job (spends credits)
Start a quoted job; this spends the user's Loopcut credits, at most the quote. Only call it after the user agreed to the quoted price or set a budget that covers it. Calling it again for the same quote returns the same job. Waits for the job and downloads the results when it finishes in time.
out_dirquote_idrequiredwait_secondsconfirm_tokenJob status and results
Wait for one job (job_id) or several (job_ids, up to 10) until they end or wait_seconds pass, then download the finished results (hash-checked; finished files are kept). A finished auto_rig job is also put into its .blend here. Running jobs report about how long is left: do other work first, then call again. A first render can spend minutes starting a GPU.
job_idjob_idsout_dirwait_secondsCancel a job
Cancel a queued or running job. Unspent credits are refunded.
job_idrequired90% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
/plugin install loopcut --marketplace loopcut-ai/loopcutclaude mcp add loopcut -- uvx loopcut mcpcodex plugin marketplace add loopcut-ai/loopcutcodex plugin add loopcut@loopcutcodex mcp add loopcut -- uvx loopcut mcp{ "mcpServers": { "loopcut": { "command": "uvx", "args": ["loopcut", "mcp"] } }}loopcut login # once per machineloopcut account # plan, credits, settingsloopcut tools # the open tools and their inputsloopcut render shot.blend --frames 1-120 --samples 256 --max-credits 300loopcut run asset_search --input '{"query": "oak floor", "kind": "material"}'loopcut run image_to_3d --input '{"prompt": "a brass desk lamp", "quality": "draft"}' --file image=lamp.png --max-credits 30loopcut quote sound_effect --input '{"prompt": "a heavy wooden door slams", "duration_seconds": 2}'loopcut run motion_from_video --file video=take.mp4 --input '{"static_camera": true}' --blend shot.blend --max-credits 60loopcut run motion_from_text --input '{"prompt": "a person jumps over a puddle", "duration_s": 4}' --yesloopcut motion-apply --blend shot.blend --bvh loopcut_outputs/JOB_ID/motion.bvh --armature Hero --in-placeloopcut motion-apply --bvh loopcut_outputs/JOB_ID --bpy # Python for a Blender that's already openloopcut start QUOTE_ID | status JOB_ID --wait | fetch JOB_ID --out DIR | cancel JOB_IDloopcut settings --private-mode off # or on; also --no-generation on|offPYTHONPATH=src python3 -m unittest discover -s tests # MCP protocol and unit testsLOOPCUT_TEST_BLENDER=/path/to/blender PYTHONPATH=src python3 -m unittest tests.test_blender_packpython3 scripts/snapshot_catalog.py # refresh the fallback catalog before a release