Hermes Agent
Hermes Agent image generation: providers, Tool Gateway and ComfyUI
Updated 19 November 2026
Hermes Agent can make pictures from a text description, and can edit a picture you give it. It does this with a built-in tool, image_generate, which sends the request to an image provider you choose. There is a second, separate route: an optional ComfyUI skill that can use a local installation or Comfy Cloud.
Facts come from the Hermes Agent 0.23.2 documentation: the image generation page, the ComfyUI skill page and the Tool Gateway page, all read on 19 November 2026.
Three routes to a picture
The docs describe the built-in tool as generating images from text prompts, with the provider and model stored in your config. Which route you take decides what you set up and who bills you.
| Route | What you need | Who bills you |
|---|---|---|
| Built-in tool with a provider you pick (FAL, OpenAI, OpenRouter and others) | An account and key with that provider, set through hermes tools | The provider, at its own prices |
| Nous Tool Gateway | A paid Nous Portal subscription, or an account that has the free tool pool | Nous, against your subscription (or the free pool) |
| ComfyUI skill, run locally | An install of ComfyUI on hardware that meets its requirements | No provider; the skill page calls local “free” and the cost is your own machine |
| ComfyUI skill, Comfy Cloud | A Comfy Cloud account and API key | Comfy, on a paid subscription |
The first two routes use the same image_generate tool and differ only in the stored provider. The ComfyUI skill is a different mechanism: a skill with scripts that drives a ComfyUI server.
Providers and models in the docs
The image generation page says eleven FAL models are supported out of the box, each with a different speed, quality and cost, and that the default is fal-ai/flux-2/klein/9b. A few rows from its table, with prices as the page gave them on 19 November 2026:
| Model | Strength, per the docs | Price listed |
|---|---|---|
fal-ai/flux-2/klein/9b (default) | Fast, crisp text | $0.006/MP |
fal-ai/flux-2-pro | Studio photorealism | $0.03/MP |
fal-ai/z-image/turbo | Bilingual English and Chinese | $0.005/MP |
fal-ai/ideogram/v3 | Typography | $0.03 to $0.09 per image |
fal-ai/recraft/v4/pro/text-to-image | Design and brand systems | $0.25 per image |
The docs say these are FAL’s prices at the time of writing and point you to fal.ai for current numbers, so treat the table as a guide to relative cost, not a quote. The full list also includes GPT Image models, Nano Banana Pro, Qwen Image and two Krea 2 models.
Beyond FAL, the same page documents these provider values for image_gen.provider:
- OpenAI. GPT Image 2.5 Flare and Sunburst, using
OPENAI_API_KEY. The docs say this is paid API usage, separate from a ChatGPT or Codex subscription. A separate OpenAI (Codex auth) provider exists and does not offer 2.5. - OpenRouter. The model picker lists OpenRouter’s live image catalogue, which the page says is more than 40 ids from the dedicated Image API plus chat-completions image models.
- Meta Model API.
meta-aiwithmuse-image-1.0, listed at $0.01 per image, text-to-image only for now. - xAI and Krea. Listed in the editing table and in the limitations as backends needing their own credentials (xAI OAuth,
KREA_API_KEY).
The OpenAI provider can also point at any OpenAI-compatible /v1/images/generations endpoint, such as a local gateway, with base_url and key_env under image_gen.openai. For picking among chat model providers rather than image ones, see our pages on Claude and DeepSeek.
Choose and configure a provider
For FAL, sign up at fal.ai and generate an API key from your dashboard. Then run the tools command in a terminal:
hermes toolsGo to Image Generation, pick a backend (Nous Subscription or FAL.ai in the example the docs give), then pick a model from the table with the arrow keys and Enter. The choice is saved to config.yaml:
image_gen:
provider: fal
model: fal-ai/flux-2/klein/9b
max_parallel_requests: 4The docs say image_gen.provider is the single selection key: nous routes through the managed Tool Gateway, while a vendor name such as fal, openai, xai or krea goes direct with your own key. They also say the runtime always follows the stored selection, so a FAL_KEY in .env is ignored while the provider is nous. Change providers through hermes tools, not by adding or removing keys. max_parallel_requests defaults to 4.
You can also set values directly, as the docs show for GPT Image 2.5 on FAL:
hermes config set image_gen.provider fal
hermes config set image_gen.model openai/gpt-image-2.5/flare/text-to-imageFor FAL model resolution, the docs say Hermes reads image_gen.model, then falls back to the FAL_IMAGE_MODEL environment variable and finally to fal-ai/flux-2/klein/9b. On FAL, GPT Image 1.5 and 2 are pinned to medium quality; the docs suggest Klein 9B or Z-Image Turbo as cheaper options.
The Nous Tool Gateway route
The Tool Gateway page says it is included with every paid Nous Portal subscription and routes Hermes tool calls, including image generation, through infrastructure Nous already runs, so you do not need a separate FAL or OpenAI account for it. It describes the tools as pay-as-you-use, billed against your Nous subscription. See Hermes Agent with Nous Portal for the plans and keys side.
- How to turn it on. Pick Nous Subscription as the image backend in
hermes tools. The page says new installs can runhermes setup --portalto log in and turn on every gateway tool at once, and that selecting the backend runs the Portal login if you are not signed in. - Models. One picker lists the FAL catalogue, native Krea 3 and Krea 2 models, and any Nous Portal image models. The model you pick decides which gateway serves the request.
- Default model. If you leave it unset, the page says a paid subscription gets Krea 3 and the free tool pool gets FLUX 2 Klein 9B.
- Free pool. The page says some accounts are entitled to a free tool pool, a small managed-tool allowance that works without a paid subscription, and that it funds the FAL models only. Krea and Portal image models need a paid subscription. It does not say which accounts qualify.
- If a model returns an HTTP 4xx. The image page says that model is not yet proxied on the portal side, and the agent tells you to switch to FAL.ai with your own
FAL_KEYor pick another model.
For whether Hermes is free in general, our is Hermes Agent free page covers the wider cost picture.
The ComfyUI skill
The skill page describes ComfyUI as a way to generate images, video and audio through diffusion workflows, and the skill text adds 3D content. Its metadata calls it an optional skill rather than a bundled one, with this install command:
hermes skills install official/creative/comfyuiThe skill uses the official comfy-cli for setup and server lifecycle, and ComfyUI’s REST and WebSocket API, driven by scripts in the skill, to run workflows. The skill page lists uses including running a specific ComfyUI workflow file, chaining steps (text to image, upscale, face restore), ControlNet, inpainting or img2img, and managing the queue, models and custom nodes. The page lists Stable Diffusion, SDXL, Flux and SD3 among the models it targets. For how skills are found and installed in general, see Hermes Agent skills.
What the skill page states about requirements:
- It asks first. When you ask to set up ComfyUI, the skill is told to ask whether you want Comfy Cloud (hosted, an API key) or Local, before running any install or hardware check.
- Local hardware. The page lists an NVIDIA GPU with at least 6 GB of VRAM (at least 8 GB for SDXL, at least 12 GB for Flux or video), or an AMD GPU with ROCm on Linux, or an Apple Silicon Mac with at least 16 GB of unified memory (32 GB recommended). It says Intel Macs and machines with no GPU will not work and should use Cloud.
- Install paths. Comfy Cloud, ComfyUI Desktop (Windows and macOS, described as beta), the Windows portable build,
comfy-clion all platforms, or a manual install. - Platforms. The metadata lists macOS, Linux and Windows.
The skill’s setup script can check the machine and install in one go:
bash scripts/comfyui_setup.shOn cost, the page says local is free but your machine must meet the requirements above. Comfy Cloud needs an API key, and the page says a paid subscription is required to actually run workflows through the API, with the free tier read-only. It points to Comfy’s own pricing page for plans. The skill also warns that a workflow is effectively code, because custom nodes run Python, so inspect workflows from untrusted sources before you run them.
Asking for a picture and where it arrives
You ask in plain language. The docs’ own examples include “Generate an image of a serene mountain landscape with cherry blossoms” and “Create a square portrait of a wise old owl, use the typography model”. They say the tool’s schema is deliberately minimal and the agent picks up whatever you configured. Aspect ratio comes in three kinds, landscape, square and portrait.
To edit instead of create, give the agent a source image as a public URL or local path. The docs say the same tool then routes to an editing endpoint when the active model supports it, and that text-to-image-only models reject image inputs with a clear error. Upscaling only runs when the agent asks for it.
The docs list delivery by platform:
| Where | How the picture arrives |
|---|---|
| CLI | Image URL printed as a markdown link |
| Telegram | Photo message with the prompt as caption |
| Discord | Embedded in a message |
| Slack | URL unfurled by Slack |
| Media message | |
| Others | URL in plain text |
The docs add that providers return hosted URLs that expire after hours or days, and that Hermes copies them to a local cache so delivery still works. To set up those chat apps, see Hermes Agent on Telegram or WhatsApp. The ComfyUI skill is different: its scripts download outputs into an output directory you give them.
Errors and fixes
- FAL selected but no key. The docs quote the error as “image_gen is configured to use fal (set via hermes tools), but FAL_KEY is not set”, followed by a hint to run
hermes tools. It is deliberate: Hermes errors rather than quietly switching provider. SetFAL_KEYor change the provider. - A key in .env that does nothing. If
image_gen.providerisnous, aFAL_KEYis ignored. Change the provider inhermes tools. - Gateway returns 4xx for one model. That model is not proxied yet. Pick another model or use FAL with your own key.
- Image input rejected. The active model is text-to-image only. Choose an edit-capable model; the docs name
fal-ai/flux-2-pro,fal-ai/gpt-image-2and several others as having editing endpoints. - OpenAI image endpoint returns 403. For the OpenAI provider the docs say an
OPENAI_PROJECT_IDset for chat can otherwise cause a 403 on projects with a model allow-list, and that Hermes sends that header blank on image requests. - ComfyUI says the workflow is in the wrong format. The scripts expect API format. Re-export from ComfyUI with Workflow, Export (API).
- ComfyUI “class_type not found” or no server. A required custom node is missing, or the server is not running. The skill suggests
check_deps.pyandauto_fix_deps.pyfor nodes, andcomfy launch --backgroundto start a server. - Comfy Cloud 403. The skill page says several API calls return 403 on a free Comfy Cloud account; running workflows needs a paid subscription.
For a problem you cannot place, the docs describe a debug switch: set IMAGE_TOOLS_DEBUG=true and per-call details (model, parameters, timing, errors) are written to a log file under ./logs.
Image generation on ReadyHermes
On ReadyHermes your assistant runs on our servers, so you do not run hermes tools. In the workspace, Bot tools has an Images setting: “Let your assistant make pictures with a fal model you choose, paid for with your own fal API key.” The options are Default, the images your assistant came with, or a fal image model you pick with your own fal key.
- The key is checked before it is saved and is never shown again.
- The assistant restarts to apply the change, which takes about a minute or two.
- Pictures made on your own key are charged by fal, not by ReadyHermes. The assistant keeps one fal key for both images and video.
Other image providers from the Hermes docs do not have a screen on ReadyHermes and we have not tested them. The ComfyUI skill is not offered or tested either, and a ReadyHermes assistant is not a GPU machine for running image models locally. Plans start from US$3.99 a month (Lite) or US$6.99 (Pro), with US$1 of AI included every month and a 1-hour free trial. See managed Hermes Agent hosting for what is included.
Questions
How does Hermes Agent generate images?
It has a built-in tool, image_generate, that sends your text prompt to the image provider stored in config.yaml under image_gen. The Hermes docs list FAL, OpenAI, OpenRouter, the Meta Model API, xAI and Krea as backends, and the default FAL model is fal-ai/flux-2/klein/9b. You pick the provider and model with hermes tools.
Is Hermes Agent image generation free?
The docs do not call the built-in tool free. FAL, OpenAI and the other providers charge at their own prices, and the Tool Gateway is billed against a Nous subscription. The Tool Gateway page does mention a free tool pool for some accounts, covering the FAL models only. The ComfyUI skill page calls running ComfyUI locally free, provided your machine meets its hardware requirements.
What is the Nous Tool Gateway for images?
It lets image generation run through Nous infrastructure with your Nous Portal login instead of a separate FAL or OpenAI account. You pick Nous Subscription as the backend in hermes tools, which stores image_gen.provider as nous. The docs say it is included with paid Nous Portal subscriptions and billed pay-as-you-use against them.
How do I install the ComfyUI skill for Hermes Agent?
The skill page lists it as an optional skill, installed with hermes skills install official/creative/comfyui. When you ask to set up ComfyUI, the skill is instructed to ask whether you want Comfy Cloud or a local install before running any install or hardware check. For local use the page lists an NVIDIA GPU with at least 6 GB of VRAM, an AMD GPU with ROCm on Linux, or an Apple Silicon Mac with at least 16 GB of unified memory.
Can I use image generation or ComfyUI on ReadyHermes?
Images, yes, with a fal model you choose and your own fal key, set in Bot tools under Images; fal charges for those pictures. ComfyUI is not offered or tested on ReadyHermes, and other image providers from the Hermes docs have no screen there.