Enables 1-click round-trip editing in DaVinci Resolve, allowing CineFlow Studio to hand off cinematic AI video projects to Resolve for editing workflows.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@CineFlow Studio turn my screenplay into a cinematic video with character sheets"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
In-IDE AI Filmmaking Studio powered by the Model Context Protocol (MCP), ComfyUI, and DaVinci Resolve. Turn screenplays and book stories into cinematic AI videos: character turnaround sheets, scene graphs, closed-mouth lip-sync guards, and 1-click round-trip editing in DaVinci Resolve.
# Run the MCP server straight from PyPI
uvx cineflow-studio
# Or install the toolkit and set up the IDE integration
pip install cineflow-studio
cineflow-setupcineflow-setup installs the CineFlow skill into your IDE skill folders
(Cursor, Claude Code, Kimi, Cline, …), writes the MCP config with the
uvx cineflow-studio command, and prints your GPU profile.
{
"mcpServers": {
"cineflow": {
"command": "uvx",
"args": ["cineflow-studio"]
}
}
}Related MCP server: Omni-Video Studio MCP
Community (free): everything local — project store, ComfyUI workflow
builder/runner, installers for models and custom nodes. Requires your own
GPU (RTX 3090/4090 recommended) with local ComfyUI on 127.0.0.1:8188.
Cloud director (CineFlow API key): set CINEFLOW_API_KEY (and
optionally CINEFLOW_CLOUD_URL) to let the CineFlow server handle
narrative planning, cross-scene continuity, and reference policy. The tier
and entitlements are decided by the CineFlow server — this client never
grants them locally.
Follow the bundled skill: cineflow_studio/skill/cineflow/SKILL.md
(installed into your IDE skill folder by cineflow-setup). Production
invariants: English-only diffusion prompts, explicit <Subject N> reference
naming, render_name aliases instead of real character names, closed-mouth
guard for silent shots.
uvx --from git+https://github.com/MariyaLoginova/cineflow-studio cineflow-mcpProprietary; see LICENSE. License terms are under consideration.
Safely pull upstream parent repository updates, refresh skills in .kimi-code and .agents, and update DaVinci Resolve scripts.
Inputs: no named parameters.
Transfer and clone characters, locations, or props from an existing project into the current target project. Automatically copies turnaround cards, plates, and voice models into target materials while maintaining project.json continuity.
Inputs: source_project (required), target_project_id (required), names, asset_types, +1 more.
Store a CineFlow API key for cloud calls. Saving is not verification: the tier and entitlements are decided by the CineFlow server on each cloud call.
Inputs: api_key (required).
Check current billing status: community (local ComfyUI) mode and whether a CineFlow API key is configured. The tier and entitlements are decided by the CineFlow server, not by this client.
Inputs: no named parameters.
Clean up temporary latents in output/latents/ and preview cache (#608) to free SSD space.
Inputs: comfy_dir.
Check parent CineFlow repository for newly released features, models, workflow blueprints, or studio updates.
Inputs: no named parameters.
Cloud mode: upload the script to the CineFlow server, enqueue the plan job, and wait for it (bounded). Requires CINEFLOW_API_KEY.
Inputs: project_id (required), script_text (required).
Get ComfyUI environment paths (comfy_dir, models_dir, custom_nodes_dir) and active API URL for the project.
Inputs: project_id (required).
Configure and save ComfyUI directory paths and API URLs for the project (stored in settings/comfyui.json).
Inputs: project_id (required), comfy_dir, comfy_url, models_dir, +2 more.
Assemble, parameterize, save, and queue a ComfyUI workflow graph (LTX, FLUX, Qwen Vision CV, or Higgs TTS). Saves reproducible .comfy.json and sidecar .cineflow.json alongside shot media.
Inputs: shot_id (required), scene_id (required), project_id (required), params, +3 more.
Synchronize project timeline, shots, and audio directly with DaVinci Resolve (or export FCPXML 1.10). Automatically installs the 'CineFlow: Edit in ComfyUI' action script for round-trip clip editing.
Inputs: project_id (required), timeline_name, install_action.
Probe and verify internal or local LLM server (DeepSeek, Qwen 2.5, Llama 3, Mistral via vLLM, Ollama, SGLang, llama-server) status, models, and studio topology (Topology 1: Single-GPU External LLM vs Topology 2: Dual-Server Internal Cluster).
Inputs: base_url.
Render all dialogue audio lines for the project to establish exact shot timing and audio pacing.
Inputs: lines (required), project_id (required).
Download a model weight into the correct ComfyUI folder (checkpoints, unet, loras, tts) with resume support or return the curl command.
Inputs: model_name (required), mirror, comfy_dir, execute_download.
Stitch rendered video clips, dialogue tracks, and background music into the final exported MP4 timeline.
Inputs: project_id (required), duck, music_path.
Generate starting keyframes for shots. Enforces reference policy (every must be addressed in prompt).
Inputs: project_id (required), frame_specs (required).
Generate an all-in-one bash/bat script (install_cineflow_models.sh) that clones all custom nodes and downloads all weights with resume support in 1 click.
Inputs: mirror, profile, comfy_dir.
Generate a supervised ComfyUI launch script (run_comfy_cineflow.sh or .bat) preconfigured with optimal VRAM flags for the user's GPU.
Inputs: profile, comfy_dir.
Inspect local computer hardware (OS, CPU, RAM, GPU, VRAM) and recommend the optimal ComfyUI memory profile (rtx_24gb, rtx_16gb, rtx_12gb, apple_silicon, or cloud_only).
Inputs: no named parameters.
Audit local ComfyUI installation, checking for installed vs missing model weights (HuggingFace) and custom nodes readiness percentage.
Inputs: comfy_dir.
Clone or update a ComfyUI custom node into custom_nodes/ and install its requirements.txt dependencies.
Inputs: node_name (required), git_url, comfy_dir.
Universal probe for internal/on-premise LLM servers (DeepSeek, Qwen 2.5, Llama 3, Mistral) and multi-server studio topology.
Inputs: base_url.
Register characters, props, and locations for the project and generate turnaround cards for human approval.
Inputs: project_id (required), props, locations, characters.
List models and weights with direct download links (HuggingFace, hf-mirror, ModelScope), file sizes, target directories, and curl commands.
Inputs: mirror, profile.
List all verified ComfyUI custom nodes required for video, turnaround, and speech generation with git clone URLs.
Inputs: no named parameters.
Initialize a new cinematic project in the studio database and storage directory.
Inputs: title (required), screenplay (required), style, aspect.
Read the canonical Source of Truth (project.json) manifest. Returns project stage, approved characters, active scene, and production history.
Inputs: project_id (required).
Initialize a physical project directory tree on disk. Stores original screenplay/book in source/, creates project.json as canonical Source of Truth, and provisions materials/ and scenes/ subdirectories.
Inputs: title (required), source_text (required), style, aspect, +4 more.
List all existing projects in studio storage with summaries of their characters, locations, scenes, and production stages.
Inputs: filter.
Update the canonical Source of Truth (project.json) manifest. Automatically tracks changes and reports impact warnings on downstream shots (e.g. if character appearance is changed after generation).
Inputs: updates (required), project_id (required).
Animate any still portrait or product photo into a 5-8 second cinematic video clip via LTX 2.5.
Inputs: image_path (required), prompt, duration.
Create an assembled commercial montage reel from a folder of source media assets.
Inputs: folder_path (required), prompt.
Single-shot high fidelity speech synthesis (Higgs v3 / Qwen TTS) with emotion and optional voice cloning.
Inputs: text (required), emotion, voice_sample.
Activate a specific scene for production. Provisions dedicated scenes/scene_XX/ directory with frames/, audio/, and video/ subfolders, and updates project.json stage.
Inputs: scene_id (required), project_id (required), scene_title.
Set director explication for a scene: emotional arc, camera style, color palette, lighting setup, beat sheet, and continuity.
Inputs: scene_id (required), project_id (required), shots, slugline, +3 more.
Retrieve the Directed Scene Graph, including nodes, narrative edges/cuts, continuity carryover, and timeline sequence.
Inputs: project_id (required).
Update the Directed Scene Graph: nodes, transitions (CUT, MATCH_CUT, CROSSFADE, L_CUT, J_CUT, SMASH_CUT), and timeline sequence.
Inputs: project_id (required), edges, nodes, timeline_sequence.
Record a generated shot artifact, saving the exact ComfyUI workflow JSON and creating a sidecar .cineflow.json alongside the media.
Inputs: shot_id (required), scene_id (required), project_id (required), media_filename (required), +8 more.
Inspect system health, ComfyUI backend reachability, active GPU jobs, and project status.
Inputs: project_id.
Render motion video clips for shots via MiniMax H3 / LTX 2.5. Enforces closed-mouth guard on non-dialogue clips (#600).
Inputs: clip_specs (required), project_id (required).
Inspect and evaluate keyframes/takes via Qwen 3.8 Vision co-located inside ComfyUI: aesthetic quality, artifact/defect detection, lip-sync closed-mouth guard (Issue #600), and English reference auto-captioning without VRAM conflicts.
Inputs: image_path (required), task, comfy_url, check_closed_mouth, +1 more.
Retrieve the exact ComfyUI JSON node graph blueprint for a specific workflow template.
Inputs: workflow_id (required).
Catalog of verified ComfyUI workflow templates (LTX-Video img2vid, FLUX turnaround, Higgs TTS, MiniMax H3).
Inputs: no named parameters.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 43 tools
Several pairs have unclear boundaries: cineflow_project_create vs cineflow_project_init, cineflow_quick_tts vs cineflow_dialogue_tts, cineflow_nodes_catalog vs cineflow_install_node, cineflow_models_catalog vs cineflow_download_model, and cineflow_workflows_catalog vs cineflow_workflow_get vs cineflow_comfy_workflow_execute. Status probing is especially fragmented across cineflow_status, cineflow_billing_status, cineflow_internal_llm_status, cineflow_deepseek_status, and cineflow_hardware_profile, where an agent could easily pick the wrong probe.
All tools use a cineflow_ prefix and snake_case, but conventions mix: verb_noun (cineflow_install_node, cineflow_download_model), noun_verb (cineflow_comfy_config_get, cineflow_workflow_get), and noun-only (cineflow_status, cineflow_auth). Ad-hoc prefixes like quick_ (quick_reel, quick_tts, quick_photo_animate) create a third naming family that doesn't map cleanly onto the rest of the lifecycle verbs.
43 tools is far above a comfortable surface for one server and several are redundant (multiple status probes, duplicate catalog/download/config get-set families). The scope spans production pipeline, infrastructure install, dependency management, and billing/auth, which is too broad to navigate reliably as one toolset.
The creative lifecycle is broadly covered: project state, scenes, scene graph, materials, frames, video render, TTS, export, plus infra and install tooling. However there are notable gaps in teardown/cleanup of projects, scenes, or materials (no delete/remove operations), leaving the CRUD surface one-sided.
Screenplay and storyboard in your AI assistant: import scripts, edit scenes, build shots.
Generate, extend and edit AI images and videos, then assemble clips on a timeline and render.
Create AI images, clips, voice-overs, music and editable storyboard films from your AI assistant.
AI film lab for filmmakers: generate and review images/clips with input provenance, cost preflight
