Sora 2 MCP Server: What Happens After the Sora Shutdown

On September 24, 2026, the Sora API was scheduled to go dark, and that date is already behind us. If a Sora 2 MCP server sits in your Claude Desktop or Cursor config, every video request it sends now lands on an endpoint OpenAI retired. The server still starts, your client still lists its tools, and then each generation call fails. No patch can fix that, because the service the server wraps no longer exists.

This article sorts out what actually ended, what you can still save, how the community MCP servers were built, and which video models take over the same job. It finishes with a working replacement setup, including a step by step run through Seedance 2.5 Lite on PicassoIA. Dates come from OpenAI's discontinuation notice and the reporting around it, so check your own account for the current state of exports.

What Actually Shut Down

The Two Dates That Matter

OpenAI retired Sora in two steps: the consumer side first, the developer side five months later.

The gap explains why so many people are searching for a "Sora 2 MCP server" only now. The April shutdown never touched API users, so automations, agents and MCP servers kept rendering clips all summer. They worked right up to the second date, and then they stopped all at once.

What Stops Working on the API

Every Sora API endpoint ended on September 24. That includes every route under the Videos API, such as content retrieval at GET /v1/videos/{video_id}/content, along with thumbnails and spritesheets. Both models, sora-2 and sora-2-pro, stopped responding.

Three details from the notice and the reporting matter for planning:

• No grace period. None was announced, and OpenAI named no official replacement endpoint.
• Conditional exports. Any final export window is conditional, and the notice says Sora data will eventually be permanently deleted. Do not plan around a last minute extension.
• Credits. There is no automatic refund promise for unused credits. The notice points to the standard ChatGPT subscription refund process.

💡 Quick check: if every call to a /v1/videos route fails while the rest of your OpenAI usage works, the shutdown is the cause. Rotating credentials or reinstalling the server will not change the result.

How Sora MCP Servers Worked

Four Tools, One Dependency

After the API opened up, several developers published Sora 2 MCP servers, including Doriandarko's sora-mcp, writingmate's sora-2-mcp, ex-takashima's openai-sora2-mcp-server and yomrishon-mcp-sora. They differ in extras: writingmate's adds FFmpeg video merging and fade animations, while ex-takashima's adds batch processing and cost estimates. At the core they expose a small set of tools. The openai-sora2-mcp-server README, for example, lists four:

• generate_video: text or image to video
• remix_video: change an existing clip with a new prompt
• get_video_status: check progress on a render
• list_videos: fetch earlier generations

Setup needed Node.js 18 or newer, an OpenAI credential stored in an environment variable, and usage Tier 2 or higher (one README mentions a minimum $10 credit purchase). The same README quoted roughly $0.40 to $1.20 per sora-2 clip and $3.00 to $12.50 per sora-2-pro clip.

Where the Servers Broke

Every one of these servers is a thin wrapper. The MCP layer, meaning tool definitions, schemas and transport, still works exactly as before. The last hop, an HTTPS call to api.openai.com/v1/videos, is what fails. That makes the breakage feel strange: the client shows the server as connected, the tools appear in the list, and only the generation call dies.

To find every affected install, search your client configs (Claude Desktop, Cursor, VS Code, Claude Code) for entries that launch an npx or node command pointing at a Sora package. Remove them. A dead server adds clutter to the tool list, and a model may call it and waste a turn on an error.

Export What You Can Still Save

If your account still lets you export, do it today. OpenAI pointed users to the Sora sunset page for exports, and its notice states that data will be permanently deleted once the final window closes, whatever that window turns out to be.

Save Prompts Before Videos

A finished clip is a short file. The prompt, the reference image and the settings behind it are the real asset, because they can be adapted for any other model. Log each generation in a spreadsheet or JSON file with these fields:

• Prompt text, word for word
• Source image, if there was one
• Model, duration and resolution
• Date and the project it belonged to
• Whether the clip shipped or stayed a draft

Build Your Own Archive

Download every MP4, thumbnail and reference image, then copy the set to a second location. Provider links are short lived, so a clip you "have" as a URL is a clip you may not have tomorrow. Cloud object storage, a NAS or a plain external drive all work. Name files by date and project, not by whatever ID the tool assigned.

💡 Rule for every tool you adopt next: write its output to storage you control on the day it is created. Vendors change plans, and Sora just showed how fast.

Replacements Worth Testing

Video Models Compared

Sora 2 was one option among many, and the field kept moving while it ran. Migration write-ups published after the announcement keep landing on the same families: Google Veo, Runway, Kling and Luma. PicassoIA's text to video category lists more than 100 entries, including all four families, so you can run one prompt through several models without opening several accounts. The table below sticks to what the catalog states.

Picking by Job

• Social clips and fast tests: Seedance 2.5 Lite or Picasso IA Video, since you can iterate without watching a counter.
• Ads and hero shots: Veo 3.1 or Ray 3.2, where polish matters more than speed.
• Cinematic sequences: Gen 4.5, Kling v3 Video or Wan 3.
• Photo animation with sound: Grok Imagine Video 1.5, listed as image to video with audio.
• Longer takes: Seedance 2.5.

Run one prompt through three candidates before you commit. A single side by side test costs an afternoon and saves weeks of switching later.

Where the Old Sora Models Stand

PicassoIA's catalog still carries pages for Sora 2 and Sora 2 Pro. Before building anything new on either, open the page and confirm they are available to you, since OpenAI's own API has reached its end date. For anything you plan to ship, choose a model that is clearly alive.

Rebuilding the MCP Workflow

The PicassoIA Connector Tools

PicassoIA offers a connector for claude.ai that exposes its GPU models as MCP tools. Generation tools return a prediction ID as soon as a GPU accepts the job, and you poll for the result until it succeeds or fails. The shape matches the old Sora flow, so the mapping is direct:

The connector's own description says generations are free on the Infinite and Wonder plans. Check the pricing page for the rest of the plan terms before you wire anything into production.

Developers who prefer plain HTTP can use the PicassoIA API at https://api.picassoia.com/v1. It follows the Replicate style: create a prediction, poll GET /v1/predictions/{id}, then fetch the output. Auth is a bearer token that starts with pia_sk_. Four models are exposed through the API and the connector: picassoia-image, picassoia-image-editor-pro, picassoia-video and seedance-2.5-lite. Limits to plan around:

• 5 concurrent predictions per account, shared across tokens and MCP connections
• 4,000 characters per prompt
• 10 MB request body
• A 3 hour timeout per prediction

Image First, Then Video

Sora's remix_video has no one to one replacement, but you can get tighter control another way: fix the first frame. Generate stills with Picasso IA Image, refine them with Picasso IA Image Editor Pro, then animate the frame you like. Both Seedance 2.5 Lite and Picasso IA Video take an input image as the opening shot and inherit its aspect ratio. Seedance 2.5 Lite also accepts an optional last frame image, so you can define where a clip begins and ends instead of remixing it afterward.

Rewrite Prompts With an LLM

Sora prompts tend to be long shot lists. The PicassoIA video models ask for cinematic, chronological descriptions: who or what is in frame, what moves, how the camera moves, what the light does. A language model handles the translation in seconds. The large language models category includes Claude Sonnet 5, GPT 5.6 Sol and Gemini 3.5 Flash. Paste in the old prompt with an instruction like this:

Rewrite this prompt as one chronological description of a 5 second clip. Start with the subject and starting pose, then the motion, then the camera movement, then the lighting. Remove anything that depends on a feature the new model lacks.

💡 Keep a short list of prompts that worked in Sora. They are the cheapest test set you will ever get for judging the replacement models.

Use Seedance 2.5 Lite on PicassoIA

Seedance 2.5 Lite is the practical first stop. It makes clips of 5 or 10 seconds with synchronized audio, renders at 480p or 720p, and Wonder members can run it without limits, so testing prompt variations costs nothing but time.

1. Open the model page and choose your input: text only, or text plus an image.
2. Write the prompt as a chronological description: starting pose, motion, camera movement, lighting.
3. Add an image if you want control of the opening frame. The clip inherits its aspect ratio. For text only, pick 16:9, 9:16, 1:1 or another ratio.
4. Set duration and resolution. Choose 5 or 10 seconds. Keep 720p for the final render, and use 480p when you only want a fast test.
5. Decide on audio. Synchronized audio is on by default. Turn it off for a silent clip.
6. Lock a seed when you want to reproduce a take, or leave it empty for a fresh result each run.
7. Generate, review and download, then copy the file into your own archive.

Parameter Settings That Matter

💡 Tip: one action per clip. A 5 second clip fits one idea, such as a head turn, a camera push or a wave. Stacking three actions makes the motion muddy.

Migration Checklist

1. Export clips, thumbnails and reference images if your account still allows it.
2. Pull every prompt and setting into a spreadsheet.
3. Remove the Sora MCP server entry from your client config so it stops failing.
4. Pick two or three replacement models from the table and run the same five prompts through each.
5. Rewrite the prompts with an LLM for the new model's style.
6. Wire up the new generation tools and test status polling.
7. Write every output to your own storage on creation.

Three Mistakes to Skip

• Swapping the model name and keeping the prompt. Sora style shot lists usually need trimming into one chronological description.
• Skipping the archive step. The Sora shutdown is the reason this step exists.
• Ignoring concurrency limits. The PicassoIA API allows 5 concurrent predictions per account, shared across tokens and connections. Queue your jobs instead of firing 20 at once.

Start Making Your Own Videos

Sora ending does not end AI video. It ends one dependency. Open Picasso IA, generate a first frame with Picasso IA Image, animate it with Seedance 2.5 Lite, then run the same idea through a second model to compare. The experiment takes minutes, and the result is a workflow that is not tied to one model. Browse the full catalog at picassoia.com/en/all-models and start with an image of your own.