← Blog

Hermes Agent Video Generation: 3 Ways and a Real Test

Hermes Agent can make videos in three ways: its built-in video_generate tool creates new clips from a prompt or an image, code-render skills turn HTML or animation code into an MP4, and a hosted video skill can change a video you already have. We tested the third way on October 15, 2026, with Riffkit, our own video tool. Hermes installed the Riffkit skill from one URL, signed in through a browser link, priced the job, and changed the outfits in our 12-second clip for 650 Riffkit credits and about $0.22 of model usage, with the finished video back about 6 minutes after it submitted. Below is what each way can do, the exact commands, what Hermes did at each step, and what changed in the video.

The car scene side by side: on the left, the original's gold satin gown and long cream gloves; on the right, the version Hermes made, with a green velvet gown and long black gloves on the same woman in the same pose

Click for the silent side-by-side: the original on the left, the version Hermes made on the right. Swap on Seedance 2.0 Fast, 480p, 12 seconds, 650 credits, about 6 minutes from submit to finished video.

Three ways to make videos with Hermes Agent

Way What it makes Starts from Good for
Built-in video_generate tool (turned off by default) A new AI clip A prompt, an image, or reference images A shot that doesn't exist yet
Code-render skills (HyperFrames, Manim, p5.js) A rendered MP4 from HTML or animation code A script or a web page Explainers, launch videos, motion graphics
A hosted video skill such as Riffkit (ours) Your existing video, re-rendered with the changes you name, or a winning video remade with your product Your own clip, a TikTok link, or a template Changing the outfit, person, setting or product in a clip that already works

The built-in tool doesn't take your own video as input. It sits in the video_gen toolset, which is off until you turn it on in hermes tools and pick a provider. Hermes' developer docs list xAI, FAL, OpenRouter and DeepInfra as its video providers, with a prompt, an image_url and reference images as inputs, and say: "Edit and extend are intentionally out of scope." So when the job is "this clip, but with a different outfit", the built-in tool would have to make a new clip from scratch.

Code-render skills draw the video themselves. The official HyperFrames skill renders MP4 or WebM from HTML, and Hermes ships Manim and p5.js skills for animation. They run on your own computer and suit text, charts, captions and product screens, not changing the people in a clip.

A hosted skill hands the work to a video service. Hermes reads the skill, runs its commands in the terminal, and the service renders the video on its own servers. That is the way we tested.

The test: change the outfits in our own clip, from Hermes

The source is our own 12-second vertical clip, made with AI and saved as a template in our Riffkit account. It cuts from a fire escape to a car to a red carpet, with the caption "can't tonight, way too much going on". Swap, the Riffkit mode that re-renders a clip with the changes you name, keeps the source's people and sound, which is why we used a clip of our own.

Setup took four commands. We installed Hermes Agent v0.22.1 with the official installer, added an OpenRouter key, picked a model, and installed the skill straight from its URL:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes config set OPENROUTER_API_KEY <your-openrouter-key>
hermes config set model.default anthropic/claude-sonnet-5.5
hermes skills install https://riffkit.ai/SKILL.md

Hermes needs a model provider; we used OpenRouter with Claude Sonnet 5.5. Before installing the skill, Hermes scanned it and returned the verdict SAFE, asked us to confirm, then saved the skill file and its reference files to ~/.hermes/skills/riffkit/. We also started each chat with only the terminal and skills tools turned on (-t terminal,skills), since the job needs nothing else.

Then we asked in plain words, with the skill loaded (hermes chat -s riffkit):

I want to change the outfits in one of my own videos with Riffkit. Sign me in to Riffkit first. Then use my analyzed template [template ID] (my red-carpet clip) in Swap mode and change her clothes: on the fire escape, the grey sweatshirt becomes an oversized cream cable-knit sweater; in the car and on the red carpet, the gold satin gown becomes a deep emerald velvet gown with long black satin gloves. Keep her the same woman: her face, hair, every shot, cut, camera move and pose, and keep the on-screen caption exactly as it is. Use Seedance 2.0 Fast at 480p. Show me the plan and the price, and wait for my go-ahead before you submit.

What Hermes did, step by step

  1. It installed the Riffkit CLI. Hermes checked for Node.js, ran npm i -g @riffkit/cli@latest (version 0.2.0) and used riffkit commands from then on. That step and the first sign-in attempt, which waits up to 90 seconds for an approval, took about 95 seconds together.
  2. It asked us to sign in. riffkit login --agent printed a link to a Riffkit page with a short code. The page showed the same code and an Approve button, with a warning that approving gives the agent full access to the account. We clicked Approve, told Hermes, and it finished signing in. No password or token went through the chat.
  3. It priced the job before doing anything. Hermes ran riffkit quote_remake and came back with a plan: Swap mode, our template, no saved character, so the same woman stays, Seedance 2.0 Fast at 480p, 13 billed seconds, 650 credits. It also flagged two things: the car and red-carpet shots show an ivory coat over her shoulders that our brief didn't mention, and the caption is burned into the picture, so the model is told to keep it but it can't be guaranteed. Then it asked "Submit?".
  4. We answered "Good catch. Make the long ivory coat a long black tailored coat. Then submit, and tell me when the video is done." Hermes added the coat to the brief and ran riffkit remake_video. The price didn't change, because a brief doesn't change the per-second rate.
  5. It waited for the video. Hermes polled with riffkit get_batch and reported back about 6 minutes after submitting, with a link that opens the video for 6 hours, the post caption and hashtags Riffkit wrote, and a note that it hadn't watched the video itself. Because it was a 480p draft, it told us to check the faces, the gloves and the caption before posting. It offered a 720p re-render on the same engine for 1,300 credits.

What changed in the video

Scene Original Result
Fire escape Grey off-shoulder sweatshirt, phone tucked on her shoulder Oversized cream cable-knit sweater; the phone stayed tucked on her shoulder
Car Gold satin gown, long ivory gloves Green velvet gown, long black gloves
Red carpet Gold gown, long ivory gloves, long ivory coat Green velvet gown, long black gloves, long black coat

Every garment we named was replaced, on the same woman. The cuts land where the original's do (we measured them: within one frame), and the caption reads the same. What drifted: the gown came out a muted, olive-leaning green rather than the deep emerald we asked for, and at 480p the far red-carpet shot is soft. We checked the clip frame by frame before writing this.

What it cost. Riffkit billed 650 credits: 13 seconds at 50 credits a second, the Swap rate of Seedance 2.0 Fast at 480p (the 12.04-second clip rounds up to 13 billed seconds). Hermes' own usage report estimated the model cost at about $0.22 for the whole session: about 308,000 tokens on Claude Sonnet 5.5, almost all of them cached prompt tokens.

How to do it yourself

  1. Install Hermes Agent with the official installer and set a model provider (hermes model walks you through it). Hermes requires a model with at least 64,000 tokens of context.
  2. Install the skill: hermes skills install https://riffkit.ai/SKILL.md.
  3. Start a chat with the skill: hermes chat -s riffkit, and ask for the change in plain words. Name each look and the scene it's in, and end with what should stay the same. Write What Changes, Not a Prompt has more on writing the brief.
  4. Sign in when Hermes sends the link: check that the code on the page matches, then click Approve.
  5. Read the plan and the price, answer what it flags, and say when to submit.
  6. Check the video before you post. Hermes doesn't watch it for you.

For your own clip instead of a template, give Hermes the file's path: the CLI sends it straight from your computer (riffkit remake_video --video ./clip.mp4; MP4, MOV, AVI, MKV or WebM, up to 100 MB and 45 seconds). For a video on your phone, Hermes can make an upload link to a Riffkit page where you add it. Change Outfit in Video with AI has the same outfit test on Seedance 2.5 at 720p, and Claude Code Video Skill covers the same skill in Claude Code.

Riffkit rates and the free plan

Riffkit bills per second of finished video, in credits. Swap rates:

Engine Swap credits per second
Seedance 2.0 Fast 50 at 480p, 100 at 720p
Seedance 2.0 60 at 480p, 120 at 720p, 300 at 1080p
Seedance 2.5 90 at 480p, 180 at 720p
MiniMax H3 80 at 768P, 160 at 2K

The plan shows the price before anything is submitted, and failed renders aren't charged.

Signup is free with no card. A free account can Swap on Seedance 2.0 Fast or Seedance 2.0 at 480p, a short clip comes back whole, and most longer ones start with a free first section that a plan finishes. Videos made on the free credits carry a small Riffkit watermark. Plans start at $15 a month for 3,000 credits. Your Hermes model provider bills separately.

Before you start

  • Approve only a link you just asked for. The sign-in page gives the agent full access to your Riffkit account.
  • Use footage you own or have the rights to. Swap keeps the source's people and sound unless you change them, so someone else's video stays theirs.
  • Keep Hermes' tools to what the job needs. Hermes can run terminal commands on your computer; we turned on only the terminal and skills tools for this test.

FAQ

Can Hermes Agent generate videos?

Yes, in three ways. Its built-in video_generate tool, in the video_gen toolset that is off by default, makes new clips from a prompt, an image or reference images through providers such as xAI, FAL, OpenRouter and DeepInfra. Code-render skills such as the official HyperFrames skill render MP4 or WebM from HTML, and bundled skills like Manim and p5.js make animations. A hosted video skill such as Riffkit can work on a video you already have, for example changing the outfit, person or setting while keeping the shots. Hermes' docs say editing and extending are out of scope for the built-in tool.

How do I turn on video generation in Hermes Agent?

Video generation is an opt-in toolset in Hermes Agent. Run hermes tools and turn on Video Generation, which also walks you through choosing a provider such as xAI, FAL, OpenRouter or DeepInfra; you can also add video_gen to --toolsets when you start hermes chat. The video_generate tool then makes a clip from a text prompt, or animates an image passed as image_url. It doesn't change a video you already have; for that, Riffkit's test used a hosted video skill installed with hermes skills install https://riffkit.ai/SKILL.md.

How do I edit an existing video with Hermes Agent?

Use a hosted video skill, because Hermes' built-in video tool doesn't take your own video as input. With Riffkit: run hermes skills install https://riffkit.ai/SKILL.md, start a chat with hermes chat -s riffkit, approve the sign-in link it sends, and ask for the change in plain words, naming each look and the scene it's in. Hermes shows the plan and price before it submits. In Riffkit's test, Hermes changed the outfits in all three scenes of a 12-second clip for 650 credits on Seedance 2.0 Fast at 480p, and the video was ready about 6 minutes after it submitted.

How do I install a video skill in Hermes Agent?

Run hermes skills install with the skill's URL, for example hermes skills install https://riffkit.ai/SKILL.md. Hermes scans the skill before installing it (Riffkit's skill came back SAFE), asks you to confirm, and stores the skill and its reference files under ~/.hermes/skills/. Load it in a chat with hermes chat -s riffkit. The Riffkit skill then installs the Riffkit CLI itself if Node.js 20 or newer is present.

How much does it cost to make a video with Hermes Agent?

Two costs add up: the model provider Hermes runs on, and the video service. In Riffkit's test, Hermes used about 308,000 tokens on Claude Sonnet 5.5 through OpenRouter, almost all of them cached prompt tokens, which its usage report estimated at about $0.22. The video itself, a 12-second Swap on Seedance 2.0 Fast at 480p, cost 650 Riffkit credits. Riffkit plans start at $15 a month for 3,000 credits, and signup is free.

Which model should I use with Hermes Agent for video tasks?

A model that follows a long skill file and runs commands reliably. Riffkit's test used Claude Sonnet 5.5 through OpenRouter: it read the skill, installed the CLI, priced the job, flagged an outfit detail the brief left out, and waited for the go-ahead before spending credits, for about $0.22 in model usage. Hermes requires a model with at least 64,000 tokens of context.

Is it safe to let Hermes Agent sign in to Riffkit?

Sign-in uses a link instead of a password: Hermes prints a link to a Riffkit page with a short code, and you approve it in your browser if the code matches. Approving gives that agent full access to your Riffkit account, so approve only a link you just asked for. Hermes shows the price and asks before it submits a video, and you can limit Hermes to the terminal and skills tools when you start the chat.

Keep reading

Higgsfield Katana Alternative: Reshoot a Video That Worked

Looking for a Higgsfield Katana alternative? What Katana is good at, three jobs that sound alike, and how to reshoot a video that already worked, in an app or in Claude.

How to Use Higgsfield Katana in Claude (2026 Guide)

How to use Higgsfield Katana in Claude: connector setup, /katana presets, what to upload, how credits work, and what Higgsfield's own pages leave unanswered.

How to Change the Background of an Existing Video with AI

Change the background of an existing video with no green screen. Three methods compared, how to choose, and our 8-second AI test: what held, what changed.