Library

OpenAI API spends chat tokens on scripts and topic batches. LM Studio uses the local OpenAI-compatible server (no OpenAI cloud). ChatGPT and Claude write natively through Desktop MCP (save_script, create_topic) — unless the writer is LM Studio, in which case they should call generate_script_via_api.

Optional: when text_provider is chatgpt/claude, scheduled jobs can draft a tagged script via OpenAI or LM Studio. Agent can still override with save_script. Does not auto-publish by itself — pipeline continues only if Hands-off / scheduler would anyway.

Default URL is http://127.0.0.1:1234/v1. Leave model blank to use the first loaded model. If generation fails, start the LM Studio local server.

When covers are manual or chatgpt, line art can still use Flux — Studio will not spend fal on title-cards. Use MCP refresh_covers to upload + re-render in one step.

No API workflow uploaded.

Use ComfyUI Save (API Format) — the /prompt graph, not the editor workflow. Studio stores it in user_data/comfyui_workflow.json. fal key can also come from env FAL_KEY or FAL_API_KEY.

Local is Resemble’s open-source Chatterbox on this machine — no OpenAI TTS charges. Install with pip install chatterbox-tts (or pip install -r requirements-tts-local.txt) and restart Studio.

Cover fills the frame; billboard is a 16:9 TV screen. Bubble-head size scales around the feet so the figure stays bottom-anchored. Both writes two mp4s (16:9 then 9:16).

Stickman head color

Recolors the yellow head fill on every file in poses/. Shadow is auto-derived (H - 10.4°, L - 8.4). Outlines stay black.

Apply to all poses Reset yellow

Only applies while Hands-off is on. Then Studio checks about every 30 seconds and starts the next due queued topic — one pipeline at a time. With Hands-off off, use Run now.

Hands-off generates topics when the queue is low (openai / LM Studio only), auto-schedules drafts, runs the pipeline, and uploads each job to YouTube as private without changing your default privacy. 0 hours = due now (FIFO).

PIN: …

How HTTP /mcp authenticates (pick one): (1) Ngrok Basic only — username/password from Settings → Ngrok; do not also send Authorization: Bearer (many clients drop Basic when both are set). (2) MCP PIN — ?mcp_pin=… or header X-MCP-Pin (optional even with ngrok). (3) Studio JWT — Authorization: Bearer <jwt> after login (per-member). PIN is hashed after save. Stdio Claude/Codex does not need HTTP auth.

Remote HTTP clients use the Public MCP URL (production domain via PUBLIC_BASE_URL, or ngrok when running). Local LAN: loopback HTTP endpoint. After a Studio update: fully quit ChatGPT, reopen, type /mcp, start a new thread.

Claude Desktop: …

Claude Code: …

Install writes mcpServers.lazykh into Claude Desktop’s config (and Claude Code if ~/.claude.json already exists). Quit and reopen Claude afterward.

Basic auth: …

Leave password blank to keep the current one. A strong password is created on first tunnel start. Give remote MCP clients only this username + password (HTTP Basic) — that is enough for admin MCP over the tunnel. Optional: still add ?mcp_pin=… if you prefer PIN auth.

Final command: ngrok http 7878 --url <your-ngrok-url> --traffic-policy-file …. Stop only ends the process Studio started.

One-time agent setup: install the ngrok app, then run ngrok config add-authtoken YOUR_TOKEN with a token from dashboard.ngrok.com (token stays in your ngrok config, not this repo). If Start fails with ERR_NGROK_107, the token was revoked — paste a fresh one and retry. After rotating the Basic password, click Start again so the traffic policy reloads.

Spend today: …

Loading…

MCP: start_gentle / ensure_gentle and stop_gentle. Stop ends Docker/local processes Studio owns; it does not quit Gentle.app. Optional: pip install pronouncing for a fuller dictionary.