Touchdown AI Video Studio

Standalone local product surface for launch videos, TikToks, ads, microdramas, and short films. It has no InferGuard dependency.

HTML brief
  -> standalone Studio gateway
  -> fixture or TrueForge named agent
  -> storyboard events
  -> zero-cost render contract
  -> hash-only receipt

Visual system

The Studio uses the Touchdown dashboard sibling system: Inter, JetBrains Mono for evidence labels, the #0066FF accent, semantic run states, 6/10/14 px radii, and one restrained panel shadow. Marketing and editorial square-card rules do not apply to this operator surface.

Run now

python3 server.py

Open http://127.0.0.1:8788. The default fixture generates a deterministic six-shot structure and a zero-cost receipt. It does not call a model or produce video pixels.

Storage

The backend automatically stores each project, storyboard, render plan, and receipt in data/studio.sqlite3. SQLite is built into Python; there is no external database service or dependency. Generated video files remain with the video provider and the project stores only their URL and receipt metadata.

GET /api/projects
GET /api/projects/<project-id>

Set TD_STUDIO_DB_PATH to move the database. Raw model prompts and credentials are never written to it.

Connect Higgsfield

higgsfield auth login
higgsfield account status

The UI supports authenticated seedance_2_0 and Higgsfield's hosted minimax_h3. It first calls higgsfield generate cost, shows the exact credit estimate and immutable job hash, and submits only after the operator confirms that exact plan. Seedance 2.0 is the default general video model. No credential is copied into the Studio.

Connect TrueForge

Start TrueForge separately, create one named agent per planning provider, then:

export TRUEFORGE_BASE_URL=http://127.0.0.1:8790
export TD_STUDIO_TRUEFORGE_AGENTS='{"glm_zai_api":"touchdown-video-glm"}'
python3 server.py

The map is the switch: each UI planning-provider value resolves to one actual TrueForge agent. The agent owns its model credentials and tools. The Studio owns the brief, acceptance contract, provider choice, render gate, and receipt.

codex_subscription is not automatically available inside TrueForge. Map it only after a harness adapter with a verified Codex OAuth path exists. Prime Agent ACP and DeepSeek Harness ACP remain visible but disabled until those adapters are implemented and tested.

TrueForge and Qodo

They serve different paths:

Video request -> Studio -> TrueForge named agent -> planning model -> storyboard
Pull request  -> GitHub CI -> Qodo review -> human merge decision

TrueForge is the runtime harness. It owns the planning-model session, tools, and context. The Studio owns the brief, approval boundary, SQLite project, and render receipt. The adapter is built and offline-tested; a live test requires a TrueForge server on port 8790 with a saved agent named touchdown-video-glm.

Qodo is code review only. This repository includes .pr_agent.toml, but Qodo does not run until its GitHub App has access to this private repository. After installation, open a pull request; CI runs the seven local tests and Qodo reviews the diff. Qodo never approves paid renders or changes runtime behavior.

Connect an authorized MiniMax H3 endpoint

The official MiniMaxAI/MiniMax-H3 repository exposes the SGLang contract:

POST /v1/videos
GET  /v1/videos/{id}
GET  /v1/videos/{id}/content

Configure an already deployed endpoint with:

export MINIMAX_H3_BASE_URL=http://127.0.0.1:30010
export MINIMAX_H3_API_TOKEN='<only when the endpoint requires it>'
export TD_STUDIO_JURISDICTION=US
export MINIMAX_H3_LICENSE_AUTHORIZATION_ID='<written MiniMax authorization id>'

The open-weight license excludes the US, EU, UK, and South Korea unless MiniMax provides separate written authorization. The adapter fails closed in those regions without that id. Hugging Face authentication or paid GPU access does not replace the license authorization. The source contract was pinned to model revision 42ed227ee7df40d41602854ae760620d6eb651fe on 2026-08-29.

Hugging Face Jobs is suitable for an authorized batch validation run, not a durable low-latency product endpoint. Jobs require a paid account, a logged-in token, multi-GPU hardware, a timeout, and explicit result persistence. No GPU job has been launched by this app.

Test

python3 -m unittest discover -s tests -v