CreditsSign in
30 TOOLS · 8 CAPABILITIES

varg MCP for any AI tool

Generate video, images, voice and finished cuts straight from your prompts — in Claude, ChatGPT, Cursor or your editor. Your agent installs it in one paste.

1

Copy the varg connector URL

You’ll paste this in the next step.

2

Add it in Claude → Connectors

This opens the Add custom connector dialog — paste the URL and click Add. Free plans allow one.

Add connector in Claude
3

Sign in and start

A browser opens on the first call. Approve it, then ask for a video.

  • 98 models
  • 11 providers
  • 30 tools
  • OAuth 2.1 + PKCE with dynamic client registration — nothing to copy, nothing to leak.
  • Revoke a connection under Settings → Connected Apps; the token dies shortly after.
  • A connected app can generate media and spend credits. It cannot mint API keys, change account settings, or reach admin functions.
ONE KEY

Stop juggling provider keys

Kling, Veo, Seedance, FLUX 3, Sora, ElevenLabs, HeyGen and the rest behind one interface, with one request shape. Swapping models is a one-line change.

Kling v3fal

Faces and human motion

kling_v3from $1.33
Veo 3.1google

Cinematic, native audio

veo_3_1from $1.68
Seedance 2.5piapi

Long shots, native audio

seedance_2_5from $0.70
Flux 3fal

Frontier, synchronized audio

flux_3from $0.90
Sora 2fal

Physical realism

sora_2$1.05
LTX-2fal

Fast and cheap

ltx_2_19b_distilled$0.53
50 video28 image8 speech

Routed across bfl · elevenlabs · fal · google · groq · heygen · higgsfield · piapi · rendi · together · varg

Live from the varg catalogSee every model →
COMPOSITING

The model will not invent your interface

Upload a real screen recording. The AI generates the actor and the scene around it, and the cut is composited against your actual UI — not an approximation of it.

Prompt onlyThe model guesses your UI — warped text, buttons that do not exist
your screencast
CompositedGenerated actor, your real recording, captions burned in — one mp4
render_video
END TO END
Five steps, one server
No Topaz, no CapCut, no separate lipsync tool, and no shuffling files between them. The agent runs the chain in one sequence and hands back a finished file.
  1. 01Shotgenerate_videotext, image or video in
  2. 02Voicegenerate_speechcast by description
  3. 03Captionstranscriptionword-level timing
  4. 04Compositerender_videoclips, overlays, audio
  5. 05Finishvideo · upscalefrom $0.16
videotext-to-video, image-to-video, video-to-video, lipsync, upscale
imagegenerate, edit, upscale
speechtext-to-speech, voice search by description
musicgenerate from a prompt
transcriptionword-level timestamps for captions
ffmpegtrim, resize, slice, probe, arbitrary commands
renderTSX composition into a single mp4
pipelinemulti-step workflows
ASYNC BY DEFAULT
Fire twenty. Collect what lands.
Video takes minutes, and a client that blocks on it times out, retries, and bills you twice. Here a call returns a job id immediately — run them in parallel, poll separately, kill the bad ones before they finish burning credits.
job_01done
job_02done
job_03running
job_04done
job_05done
job_06cancelled
job_07done
job_08done
job_09done
job_10running
job_11done
job_12done
job_13cancelled
job_14done
job_15done
job_16done
job_17running
job_18done
job_19done
job_20done
15 done3 running2 cancelled — not billed to completion
COST CONTROL
See the price before you spend it
Every generation is priced from a published rule before anything runs. The agent can quote it, check the balance, and wait for a yes — instead of surfacing the number on your invoice.
youFour 10-second shots on Kling v3. What does that cost?
estimate_costkling_v3 · 10s · 16:9 · ×4
  • Price per shotquoted from the live rule
  • Balance aftercheck_balance()
RunCancel
An illustration of the estimate step. Real figures come from estimate_cost, reading the same pricing rules published on /v2/models.
LINEAGE
One good take, ten variations
Every file remembers how it was made — the job, the tool, the model and the exact input. Paste a URL back and the agent recovers the recipe, changes one thing, and re-runs it.
source fileget_lineage(url)
hook Aprompt changed
hook Bprompt changed
verticalaspect_ratio changed
new actorreference changed
TEAMS & STORAGE
Every connection picks its own team
Claude bills Marketing while Cursor bills Product — the setting is per connection and it persists. Output lands in that team’s library, in whichever folder you point it at.
MarketingClaude
current
  • q3-launch
  • paid-social
ProductCursor
  • onboarding-demo
Client — NorthwindVS Code
  • brand-films
  • cutdowns
BATTERIES INCLUDED
Templates, voices, music, captions
Start from a published template instead of a blank file, cast a voice by describing it rather than scrolling a list, and get word-level timings back for captions that actually line up.
Templates
Search published templates by tag, render one as-is, or pull its TSX source, edit it in the chat, and render your own version.
talking headproduct showcaseUGCslideshow
Voice by description
Describe the read you want and get ranked suggestions, or filter on gender, accent and language. ElevenLabs under the hood.
Music & captions
Score a cut from a prompt, and burn in captions from a transcription that carries a start and end timestamp for every word.
casting a voice
Point your agent at varg.
One URL, one approval, thirty tools.
DocsGitHub