walid@portfolio:~/lab/pexo-video-skills$
cd../lab
03ideaSep 2026

One key, every video model

Wiring a hosted video agent into your coding agent — it plans the shots, picks the model per shot, and returns a finished MP4

One install and one API key put a hosted video agent behind your coding agent: you describe a video in conversation, it writes the script, picks a model per shot, generates the clips, adds music, narration and subtitles, and hands back a single finished MP4. The skills are free and MIT; the generations cost credits, and nothing paid runs without you approving an estimate first.

VideoAgent SkillsClaude CodeMITCreditsThe repo ↗ElevenLabs ↗fal.ai ↗
i
What makes it different from a model wrapper

Most video skills wrap one model and hand back a raw clip. This one returns a finished, multi-shot video — script, shot sequencing, per-shot model selection, music, narration, subtitles and transitions, assembled. You give it a brief in conversation and take delivery of an MP4. The trade is that you are not choosing the models for the automated path; a hosted backend is. Separate studio skills exist for when you do want that control.

The pieces

Four kinds of skill, two levels of control

SkillWhat you getControl
The agent skillA finished video from a conversation — script, shots, music, subtitles, assembly.Fully automated
13 scenario skillsThe same agent with a focused entry point per job — video ad, shorts maker, product video and so on.Fully automated
The director skillStoryboard first, then images, clips and audio shot by shot.You approve the plan
Video / image / audio studiosOne clip, image or audio file from a model you name yourself.You pick the model
!
On the “10+ models” headline — checked against the repository

The repository does advertise 10+ video models, and its own README carries that badge. But it names only four of them for the auto-selecting agent — Seedance 2, Kling 3.0, HappyHorse and Grok — with the rest given as “and others”. The full list is not published. What you can actually name and pick by hand is 7 in the video studio and 8 in the image studio. Both figures were confirmed against the README today. Nothing dishonest here, but “10+” is a backend claim you cannot audit, while 7 and 8 are lists you can.

Section 2

What you need — and what you do not

One key covers every model the automated agent uses. No per-model API keys, no FFmpeg, no Python.

RequiredOptional
A coding agent — Claude Code, Cursor, Codex or OpenClaw.Node.js — only for the studio skills’ generation tools.
A terminal with npx for the one-line installer.ElevenLabs and fal.ai keys — only for the audio studio (speech, effects, music).
curl, jq and file — the bundled doctor script checks for these.Git — only if you clone the repository instead of using the installer.
An account and API key from the hosted service.A paid subscription — needed only for watermark-free downloads.
Section 3

Install it

The global flag installs to a shared skills folder and symlinks it in, so every project sees it. Drop the flag to install into the current project only. Add another --skill argument for each studio you want.

terminal6 lines
# fastest — one command, global
npx skills add https://github.com/pexoai/pexo-skills --skill pexo-agent -g

# or via the plugin marketplace, which installs all 7 core skills
/plugin marketplace add pexoai/pexo-skills
/plugin install pexo@pexo
Section 4

Give it your key

The key lives in this file and never in a prompt. The base URL line is optional, and must be exactly the service’s own address if you set it at all.

~/.pexo/config4 lines
PEXO_API_KEY="YOUR_API_KEY_HERE"

# optional, must be exactly this if set:
PEXO_BASE_URL="https://pexo.ai"
Verify before spending

Run the doctor, then a read-only test

The doctor checks the config, the required command-line tools, network reachability and whether the key actually works. Neither call spends credits — which is the point of doing them first.

terminal + prompt6 lines
bash ~/.claude/skills/pexo-agent/scripts/pexo-doctor.sh

# then, in your agent:
Use the pexo-agent skill. Run pexo-doctor.sh and pexo-project-list.sh, tell me
whether Pexo is ready, and list the video capabilities available to me.
Do not create a project or generate anything yet.
Section 5

The seven you can pick by hand

These are the video-studio models. All run through a hosted proxy, so none of them needs its own key.

ModelBest for
seedanceClips with synchronised audio, roughly 4 to 12 seconds.
klingMulti-element, character and keyframe control.
veoMultiple reference images, with a 4K option.
minimaxSubject-reference images and character consistency.
hunyuanText-to-video, and video-to-video style transfer.
pixverseStylised image-to-video.
grokVideo editing driven by a reference video.

The other two studios

StudioWhat it offersKeys
ImageEight models — Midjourney, three Flux variants, Ideogram, Recraft, SDXL and a Gemini-based one.Free tier, auto-issued token
AudioSpeech, sound effects and voice cloning, plus music generation.Your own provider keys

Driving a studio directly

Install the studio skill first. This is the escape hatch for when you want one specific clip rather than a finished film.

terminal5 lines
node ~/.claude/skills/videoagent-video-studio/tools/generate.js --list-models

node ~/.claude/skills/videoagent-video-studio/tools/generate.js \
  --mode text-to-video --model kling --aspect-ratio 9:16 --duration 5 \
  --prompt "slow push-in on a rain-soaked neon street, cinematic"
Section 6

What actually happens on a run

You only talk; the agent runs the scripts. Note where the gate sits — the estimate comes before any generation, and approval is a separate explicit call.

idea → mp413 lines
IDEA              you describe the video in your agent
   ↓
PROJECT CREATED   a project id comes back
   ↓
PLAN + MODELS     the backend plans shots and selects models;
                  it may ask questions or offer preview options
   ↓
CREDIT ESTIMATE   → YOU APPROVE. Nothing paid runs before this.
   ↓
GENERATE          clips, music, narration, subtitles, transitions
                  — your agent polls roughly every 60 seconds
   ↓
FINISHED MP4      downloaded locally, plus a project page online
i
How long it takes

The repository’s own figures: around eight minutes for a fifteen-second, three-shot ad, and around twenty minutes for a sixty-second brand video. Your agent hands you the project link so you can watch progress there rather than sitting in the terminal. Budget the wait — this is not an interactive loop.

Section 7

The master prompt

Written to match how the thing actually works: the backend selects models, so you ask for the plan and the estimate rather than trying to dictate models. The line that matters most is the one forbidding it to approve on your behalf.

master-prompt.txt17 lines
Use the Pexo video skills to create a 30-60 second vertical video about:
[VIDEO IDEA]

Format: 9:16   Platform: Reels / TikTok / Shorts
Style: fast-paced, premium, cinematic   Length: 30-60 seconds

Send this brief to Pexo as written. If Pexo asks questions or shows preview
options, relay them to me and wait for my answer.

Before any paid generation, show me Pexo's credit estimate and ask for my
approval — do not approve on my behalf.

Use music, subtitles, narration, and transitions only where Pexo supports them.
Do not publish anything anywhere.

When finished, give me the full download URL, the project link, and the local
file path where every generated file is saved.
Section 8

Three to try

Illustrative only — no brand affiliation implied. Each one keeps the estimate gate in place.

prompts.txt10 lines
1. Create a cinematic 30-second commercial for a futuristic running shoe
   from a fictional sportswear brand. 9:16. Show me the credit estimate
   before generating.

2. Create a 45-second faceless short explaining one new AI technology,
   with narration and subtitles. Ask me which technology first.

3. Create a cinematic trailer showing what a major city might look like
   in 2050. 4 shots, 16:9, moody score. Show me the plan and cost
   before generating.
Section 9

Where the output lands

WhatWhere
The finished videoDownloaded to a local temp folder — the returned JSON gives both the local path and the full URL.
The same video onlineYour project page, which keeps every version.
Preview options offered mid-runFetched per asset id into the same local folder.
Studio clips, images and audioReturned as a URL in the script’s JSON — not saved locally unless you download it.
Subtitles, music, narrationBaked into the finished MP4. Not delivered as separate files by this skill.
!
One finished film, not a folder of assets

The automated agent delivers a single rendered MP4 — if you wanted the loose clips to edit yourself, that is the studio skills, which are the exact opposite shape. Worth knowing before you plan an edit around it. Downloads are watermark-free by default, but that depends on an active subscription; without one, ask for the watermarked variant explicitly.

Section 10

What costs money

The distinction that matters: the skills are free and open source, the generations are not.

ItemFree or paid
The skills and installerFree — MIT open source, confirmed against the repository.
Doctor and project-list checksFree — no credits consumed, which is why they make a good first move.
Video studio, 7 modelsFree by default through the hosted proxy, per the skill definition.
Image studio, 8 modelsFree tier — 100 uses per auto-issued token.
Agent generationsPaid, in credits. The estimate appears at the confirm step and you approve it.
Watermark-free downloadsPaid — requires an active subscription.
Audio studioPaid — your own provider keys and their usage.
Your coding agentWhatever your existing plan costs.
Section 11

Troubleshooting

ProblemFix
Config not found, or the key is not setCreate the config file with the key line, then re-run the doctor.
401, invalid API keyExpired or mistyped. Issue a new one from your account and replace it in the config.
curl, jq or file not foundInstall them with your package manager — the doctor tells you which is missing.
429 on project createEither a project is still running, or credits are out. Read the message; it distinguishes the two.
Still waiting after thirty minutesOpen the project page. The skill stops polling and asks whether to keep waiting or start over.
Start here

Quick start

✓Install your coding agent.✓Install the video skill with the one-line installer.✓Put your API key in the config file.✓Run the doctor and confirm it reports ready — this costs nothing.✓Pick a video idea.✓Give the agent the master prompt.✓Review the shot plan and any preview options it offers.✓Approve the credit estimate deliberately, not reflexively.✓Let it generate, and collect the MP4 from the local folder and the project page.
i
Provenance

Checked against the repository on 18 September 2026 rather than taken on trust. The licence is MIT and the version file still reads 0.3.12 — the same release the source guide pinned in August, so its commands and paths are current despite ongoing repository activity. The install command matches the README exactly. The model counts were verified too: the README does carry a 10+ badge while naming only four models for the automated path, and the studio lists really are 7 video and 8 image models. A referral code that appeared in the source’s credits section is not reproduced here, and neither is the author’s byline or follow prompt.

← previous
Give an agent $100 and a shutdown switch
next →
Make one job repeatable
← all experiments