sogni

Sogni AI image generation backend for image_generate: text-to-image and image-to-image (identity edit) on Sogni's decentralised GPU network, driven through Sogni's own sogni-agent CLI. Registers as an image_gen provider, so the normal image_generate tool works unchanged — aspect ratio and reference images are threaded through. Disclosure — every generation submits a job to the Sogni network via the sogni-agent-hermes CLI, which this plugin spawns as a subprocess, so unlike Hermes's bundled providers it is not zero-install: it needs Node.js and npm install -g @sogni-ai/sogni-creative-agent-skill@3.56.1 (or SOGNI_AGENT_BIN pointing at an existing binary). The prompt and any reference image (including a local file path the model passes) are sent to api.sogni.ai on the way to the network and the finished image is downloaded from media.sogni.ai; the plugin reads SOGNI_API_KEY through Hermes's secret scope and hands the CLI a minimal environment (PATH, HOME, LANG, SOGNI_API_KEY and, if set, NODE_OPTIONS, NPM_CONFIG_PREFIX, XDG_CACHE_HOME), so no other Hermes provider keys reach it, and it presents an honest custom User-Agent because Cloudflare (Error 1010) rejects library-default ones. Generation is billed to your Sogni account per job with no Hermes approval prompt, and the plugin reports the model used but enforces no budget cap, so an unattended run can spend your balance. Generated media is written under $HERMES_HOME/cache/images/. No telemetry, no stored credentials, no background processes beyond the CLI call, no self-updates.

RichardChua1234Toolsv1.0.0Updated 1 hour ago

About

Sogni AI — image generation provider

Adds Sogni AI as an image_gen backend for Hermes, so the
agent can generate and edit images on Sogni's decentralised GPU network.

hermes plugins install sogni

Once installed and selected as the active image_gen provider, the normal tool
surface works unchanged:

image_generate(prompt="cinematic moonlit mountain lake")
image_generate(prompt="make it night; add fireflies", image_url="./lake.png")

Why this exists

Hermes ships image backends for FAL, DeepInfra, OpenAI, OpenRouter, xAI and
Krea (Krea.ai). Sogni AI is a different vendor — a decentralised GPU network
fronting its own model families plus editing, upscaling, 3D reconstruction and
video — and nothing in Hermes or the plugin catalog could reach it. This plugin
closes that gap.

Install

The Sogni CLI is a hard requirement. This plugin drives the vendor's own
supported client rather than reimplementing Sogni's asynchronous workflow
protocol:
```
npm install -g @sogni-ai/sogni-creative-agent-skill@3.56.1
sogni-agent-hermes --version
```
API key. Get one from https://dashboard.sogni.ai (account menu) and set
SOGNI_API_KEY. Hermes reads it through its secret scope.

Enable the plugin and select the provider:

hermes plugins enable sogni
hermes config set image_gen.provider sogni

If the CLI is not on PATH, set SOGNI_AGENT_BIN to its full path.

Models

Model ids below were returned by the live API
(sogni-agent-hermes --search-models), not copied from documentation. Only
verified ids are advertised — an unverified id would fail at generation time
instead of being honestly absent.

| Provider id | Sogni model id | Notes |
| --- | --- | --- |
| krea-2-turbo (default) | krea2_turbo_fp8_scaled | Fast general purpose |
| dark-beast-krea2 | dark_beast_krea2_fp8 | Higher fidelity, more workers |
| krea-2-identity-edit | krea2_identity_edit_v1_2 | Editing; needs a source image |
| dark-beast-krea2-identity-edit | dark_beast_krea2_identity_edit_v1_2 | Editing; needs a source image |

Aspect ratios map to explicit pixels: landscape 1216×832, square 1024×1024,
portrait 832×1216.

Verified behaviour

GET https://api.sogni.ai/v1/models returns 200 with an honest
User-Agent. The default Python-urllib signature is refused with Cloudflare
Error 1010 (HTTP 403) — a client-signature block, not an authentication
failure. This plugin does not spoof a browser string; if you see a 403
from Sogni, that is why a plain client struggles.
Image model ids and worker counts come from a live --search-models call.
CLI flags confirmed against --help on v3.56.1: -m/--model, -w/--width,
-h/--height, -n/--count, -c/--context, --output-format, -o/--output,
--json.

Capabilities

provides_tools:      []
provides_hooks:      []
provides_middleware: []
requires_env:
  - SOGNI_API_KEY
  - SOGNI_AGENT_BIN

Limitations

The CLI must be installed. is_available() returns False without it, so
the provider is never selected when it cannot work. This is a Node/npm
dependency outside Python's reach.
No spend cap. Sogni bills per job in capacity units and the CLI reports
cost after the fact. There is no honest upfront ceiling to enforce, so an
unattended run can spend your balance.
Only the four models listed above are advertised. Sogni fronts more (Z-Image,
FLUX, Chroma, Qwen-Image, upscalers, 3D, video) — those need their ids
verified live before being added here.
The CLI's --json schema is the vendor's to change. Result extraction walks
the common shapes and reports an honest error rather than inventing a success
if none match.

Disclosure (catalog rule 13)

What a user should know before installing:

Runs a shell command. Each generation executes the local
sogni-agent-hermes binary as a child process with the prompt and any
reference image paths as arguments. Nothing else is run, and no shell
interpreter is involved — arguments are passed as a list, never through a
shell.
Network. The CLI talks to Sogni's hosted service (api.sogni.ai) and the
GPU network: the prompt, image dimensions, model choice and any reference
image are sent to Sogni. The result is fetched back from Sogni's storage.
Credentials. SOGNI_API_KEY is read through
agent.secret_scope.get_secret, never from a .env file, auth.json or a
bare os.environ read — a profile with no key of its own will not borrow
another profile's. The key is passed to the CLI in its environment and is never
logged or written by this plugin.
Reads. Reference images you pass are read from local paths and handed to
the CLI for upload.
Billing. Generation is billed to your Sogni balance. Cost is reported, not
capped.
No telemetry, no background processes, no self-update, no stored
credentials. Zero Python dependencies — standard library only.

Files

__init__.py — the provider and register(ctx)
plugin.yaml — manifest

README from the reviewed commit 2e713ff; it updates when the author re-pins.

More by RichardChua1234

2 plugins by RichardChua1234 in the catalog.

RichardChua1234leonardoLeonardo.AI video generation backend for video_generate: Motion 2.0 text-to-video and image-to-video through the Leonardo REST API, using only the Python standard library. Disclosure — the prompt and any reference image URL are sent to cloud.leonardo.ai and the finished clip is fetched from cdn.leonardo.ai, authenticated with your LEONARDO_API_KEY, which the plugin reads through Hermes's secret scope and never from a file or the ambient environment; generation is billed to your Leonardo API token balance (about 200 tokens per 5-second clip, purchased separately from a Leonardo web subscription); the job is polled in-process with a bounded wall-clock deadline, and the token cost is reported back but no budget is enforced, so an unattended run can spend your balance; no telemetry, no shell commands, no background processes, no stored credentials and no self-updates. Zero Python dependencies.Communitytools

Developer Mode

Explore built-in Hermes toolsets and actions.

HermesAgent

Developer?

Create a Hermes extension for Desktop or agent workflows.

Developer Section