I wanted to compare AI video models side by side without opening four sites and four subscriptions. So I put them
behind one API and wrapped that API as an MCP server. Now I ask Claude Code for a clip in plain words, and the file
lands in a folder on my disk.
Disclosure: I build Turbogen, the service this post uses. The MCP server is MIT-licensed and on PyPI.
Setup (2 minutes)
You need an API key (it starts with lk_) and uv. Nothing else to install:
uvx fetches the server on first run.
Claude Code
claude mcp add turbogen --env TURBOGEN_API_KEY=lk_your_key -- uvx turbogen-mcp
Claude Desktop / Cursor (claude_desktop_config.json or .cursor/mcp.json):
{
"mcpServers": {
"turbogen": {
"command": "uvx",
"args": ["turbogen-mcp"],
"env": { "TURBOGEN_API_KEY": "lk_your_key" }
}
}
}
The server exposes seven tools: list_models, generate_image, generate_video, generate_audio,
upscale_image, get_balance and job_status.
1. Ask what is available and what it costs
"List the video models with their cheapest option and price."
list_models answers with every model, its options and the price in credits (1 credit = $0.01), for example:
- seedance-2.5 [video] Seedance 2.5
Duration: 5s=60cr, 10s=119cr, 15s=179cr, 20s=239cr, ... 5s 720p=134cr, ...
- kling-2.5 [video] Kling 2.5
Duration: 5s=18cr, 10s=36cr
- veo-3.1-fast [video] Veo 3.1 Fast
Duration: 4s=12cr, 6s=18cr, 8s=24cr
The assistant reads this before it picks a model, so "use the cheapest Kling" just works.
2. A start frame with Nano Banana Pro
"Generate a 16:9 image with nano-banana-pro: a red vintage bicycle leaning on a wall in a quiet Tokyo alley,
morning light."
Saved: ~/Downloads/turbogen/nano-banana-pro-a-red-vintage-bicycle-leaning-on-a-weath-1bf148a6.jpg
Cost: 6 credits · Credits left: 309
37 seconds from prompt to file.
3. The same frame through three video models
"Animate that image with seedance-2.5, kling-2.5 and veo-3.1-fast: slow push-in toward the bicycle, a few
cherry petals drift past, gentle morning light."
The assistant runs the three jobs in parallel.
| Model | Length | Output | Credits |
|---|---|---|---|
| Seedance 2.5 | 5 s | 856×478 | 60 |
| Kling 2.5 | 5 s | 1284×716 | 18 |
| Veo 3.1 Fast | 4 s | 1920×1080 | 12 |
Seedance 2.5
Kling 2.5
Veo 3.1 Fast
Same frame, same prompt, three different readings of it:
- Seedance 2.5 took "push-in" literally: the camera travels right up to the bicycle and the frame ends on the saddle and basket. The strongest camera move of the three, at the lowest resolution of this option.
- Kling 2.5 stayed closest to the start frame: a gentle drift forward, a few petals, nothing invented. The cheapest per second here and the safest if the frame must not change.
- Veo 3.1 Fast went for atmosphere: a full shower of petals, only a slight camera move, and the only 1080p output. Shortest clip, lowest price.
None of them is "the best". Which one wins depends on whether you need motion, fidelity or resolution, and that is
exactly why I wanted to try them side by side.
4. Check the balance
"How many credits do I have left?"
Credits: 219
The whole run (one image, three clips) cost 96 credits, about $0.96. One extra job failed on the model side
(Seedance 2.0 Mini was briefly unavailable) and its credits came back automatically.
How it works under the hood
Video jobs take from 30 seconds to a few minutes, which is longer than some MCP clients like to wait. So each
generation tool submits a job, polls every 5 seconds for up to 10 minutes, downloads the file and returns the local
path, the cost and the remaining balance. If a job runs longer, the tool returns the job id and the agent can call
job_status later. Failed jobs are refunded automatically, so a retry loop never charges twice.
One thing I learned while writing this post: three parallel jobs polling separately, plus the web app open in a
browser on the same IP, ran into the API's rate limit. Version 0.1.2 makes concurrent jobs share one status request
and wait out a 429 instead of failing the generation.
The same thing without MCP (plain REST)
import time, httpx
API = "https://turbogen.app/api/jobs"
H = {"Authorization": "Bearer lk_your_key"}
# submit: form fields (+ optional "refs" files for image-to-video)
job = httpx.post(API, headers=H, data={
"model": "veo-3.1-fast", "aspect": "16:9", "option": "4s",
"prompt": "a red vintage bicycle in a Tokyo alley, slow push-in"}).json()["job"]
# poll: the job list carries status and your remaining credits
while job["status"] in ("queued", "running"):
time.sleep(3)
jobs = httpx.get(API, headers=H, params={"limit": 50}).json()["jobs"]
job = next(j for j in jobs if j["id"] == job["id"])
print(job["status"])
# the file: GET {API}/{job_id}/file?download=true
Links
- Server on PyPI: https://pypi.org/project/turbogen-mcp/
- MCP Registry:
app.turbogen/turbogen - API docs: https://turbogen.app/api-docs
- Prices per model: https://turbogen.app/pricing
Which model would you want to see compared next?




Top comments (0)