hermes-aipg-provider
A Hermes Agent model-provider plugin for the
AI Power Grid, an OpenAI-compatible inference API whose open models are served
by independent GPU operators.
Community-written and unofficial. The author is a user of the grid, not part of the AIPG team.
Install
hermes plugins install speteai/hermes-aipg-provider
hermes plugins enable aipg-community
Or by hand: put __init__.py and plugin.yaml into $HERMES_HOME/plugins/model-providers/aipg/.
Then set your Grid API key in ~/.hermes/.env:
AIPG_API_KEY=your-key-here
and select the provider, for example in ~/.hermes/config.yaml:
model:
default: "qwen38-flash-next-nvfp4"
provider: "aipg"
What it adds
- An
aipgprovider (aliasesai-power-grid,aipowergrid) selectable withhermes model,
--provider aipg, or/model <model-id> --provider aipg. - The model list comes live from the grid's
/v1/modelsendpoint. qwen38-flash-next-nvfp4as the fallback and auxiliary model.
It is a plain provider profile: no tools, no hooks, no middleware.
Tested
On 2026-11-15, Hermes Agent v0.23.1 on Linux x64, against the live grid:
hermes plugins validate .passes.hermes plugins install speteai/hermes-aipg-providerplusenable, then a tool-using prompt through the
installed plugin, with the key in.envand with the key as an environment variable.- With
provider: "aipg"andqwen38-flash-next-nvfp4, the agent wrote a Python script, ran it and
wrote the result to a file (27 s, 4 model calls, about 52,000 tokens).
Not tested: Windows, macOS, the desktop app, long sessions.
Known limits
gpt-oss-120bis listed by the grid but does not work for agent runs at the moment: on multi-step
tasks its tool call comes back inside the reasoning text, and the run ends without a reply
(aipg-documentation#70). Use
qwen38-flash-next-nvfp4.grid/qwen3:8bhas an 8,192-token context, which is too small for Hermes.- Hermes showed the cost of the test run as unknown. The grid's prices are at
https://api.aipowergrid.io/v1/pricing; the account balance shows the real charge.
Disclosure
- Network calls: only to
https://api.aipowergrid.io/v1(inference and the/modelslist), made by
Hermes with the API key you provide. - Your prompts, tool definitions and tool results are sent to that API and processed in plain text on
computers run by independent operators. Do not use it for confidential material without thinking
about that. - Usage is billed per token from a prepaid balance of your grid account. The plugin makes no payments.
- The API key is resolved through Hermes's own provider-credential route (
AIPG_API_KEY,
optionalAIPG_BASE_URL). The plugin reads no files and stores nothing. - No telemetry, no background processes, no self-updating code, no core overrides. It registers through
the publicregister_provider()/ProviderProfileAPI only.
Why standalone
Hermes's contribution policy asks for third-party integrations to ship as plugins rather than in the
core tree (see
CONTRIBUTING.md).
License
MIT
README from the reviewed commit 88dcdc1; it updates when the author re-pins.

speteai