Higgsfield demos GPT-6 Astra controlling creative apps through MCP
Higgsfield demos show GPT-6 Astra completing editable work in desktop creative software through its MCP integration. Examples include Blender scenes, After Effects motion graphics, Figma designs, DaVinci color grading, and Illustrator work.

TL;DR
- Higgsfield’s pipeline puts scene code between a floor plan and a Cycles-rendered Blender location, as the floor-plan demo documents.
- The claimed deliverables remain native and editable: the After Effects screen recording describes reusable transparent compositions, while the Figma walkthrough says the reconstructed file stays editable.
- One MCP run combined character design, modeling, UVs, rigging, self-evaluation, retry, and final stylization in 30 minutes, according to the nine-stage desktop run.
- Desktop control included a recovery path: a frozen file picker sent the Illustrator task through Finder, according to the Illustrator export run.
A 3D camera teardown breaks into 1,025 modeled objects in a Higgsfield demo. In an audio-dubbing task, matvelloso reported that Astra patched repeated and mixed-language segments, isolated the sound effects, reinserted them, adjusted volume, and checked transcripts in matvelloso's report.
Scene code to Cycles
Higgsfield describes a handoff where Astra turns the floor plan into scene code, then its system builds the location in Blender and renders it with Cycles.
The Opera House example makes the same claim, with the landmark reconstruction saying the building came out as code before Blender produced clean, editable geometry.
Editable application artifacts
The After Effects run is framed as a 20-minute recreation of a reference video whose motion elements land in separate editable compositions with transparent backgrounds.
Its Figma counterpart rebuilds a reference in stages, outline, light and volume, face, emblem, then texture, inside an editable native file.
Repair loops
A DaVinci Resolve demo starts from flat Apple Log 2 footage and asks Astra to match a reference image. Higgsfield says the grade took four minutes.
The execution host supplies the other half of this interaction. OpenAI’s computer-use integration recipe says the host must execute requested actions and capture screenshots throughout a session; for native desktop apps, it specifies a VM or container that translates returned actions into operating-system input events.
Nine-stage MCP run
Higgsfield’s most explicit prompt makes the agent carry a character from concept to final cartoon through its MCP bridge.
- Design a character concept with Higgsfield Soul 2.0.
- Build a textured 3D model.
- Import it into Blender.
- Retopologize the mesh.
- Create a UV map.
- Build a character rig.
- Evaluate production readiness.
- Redo prior stages if the result falls short.
- Turn it into a cartoon with Seedance 2.5.
Stages seven and eight turn evaluation and retry into explicit parts of the prompt. A separate review by Matt Shumer describes a coordinator-plus-implementer setup for longer runs, while cautioning that agents can get absorbed in details and plateau without careful coordination.
Screenshots, VMs and access
The Illustrator demo offers a useful desktop edge case: it says Astra built the native document, then routed through Finder when the application’s file picker froze.
OpenAI’s launch announcement frames Astra around computer use. Its September 3 release notes described a rollout limited to selected organizations, said general availability had not yet begun, and noted that safety monitoring can pause or stop a conversation when an agent may have misread instructions.
56 million tokens
A cost tally using GPT-6 Astra Standard API pricing puts one workflow at an estimated API-equivalent $88.37 across 56,180,375 tokens.
- Cached input reads: 54,644,088 tokens and $61.89.
- Cache writes: 1,462,322 tokens and $22.60.
- Output: 68,567 tokens, including 26,175 reasoning tokens, and $3.83.
The tally says it aggregated 314 response records, including compaction, and excludes separate tool fees.