Video: "Hermes Agent V0.13 Just Changed AI Agents Forever!" by Julian Goldie on YouTube.
Pluggable providers: swap your AI model without touching the core
Previously, adding support for a new AI provider to Hermes meant modifying the core codebase — something only contributors comfortable with the internals could do. V0.13 introduces a ProviderProfile abstraction that turns provider integration into a plugin. You add a profile file, point it at the provider's API, and Hermes picks it up without any changes to the main code.
In practice this means teams can switch between OpenAI, Anthropic, Groq, or a local model like Gemma or Llama without rearchitecting their setup. It also means third parties can publish provider plugins the way you'd publish a browser extension. That is a meaningfully different model from "edit the source and recompile."
Eight P0 security fixes you should know about
A P0 fix is a critical security closure — the kind that gets shipped as a priority, not queued for the next feature release. V0.13 ships eight of them. The most significant: output redaction is now on by default, Discord role-allowlists are now scoped to individual servers rather than globally, and WhatsApp rejects messages from unknown contacts by default.
These matter because Hermes is often connected to messaging platforms and given access to files and APIs. Before v0.13, a poorly configured setup could leak responses to unexpected recipients or allow outside parties to interact with your agent. The new defaults close the most obvious gaps. Worth checking your existing config against the release notes if you have an older installation running.
Checkpoints v2: state persistence that actually works
The original checkpoint system in Hermes was basic — it saved state but had no disk guardrails, so long-running jobs could balloon storage without warning and corrupt state in edge cases. Checkpoints v2 rewrites the persistence layer with proper pruning, disk limits you can configure, and better recovery when a checkpoint is incomplete.
For teams running Hermes on shared servers or inside CI pipelines, this matters more than it sounds. An agent that fills a disk mid-run, or that fails to resume cleanly after a crash, is an operational problem. The v2 persistence is considerably more reliable in those scenarios.
What the multimodal additions actually add
V0.13 also ships a video_analyze tool — an agent can now pass a video clip to a compatible multimodal model (Gemini and a few others) and get a structured analysis back. Combined with the xAI Custom Voices TTS provider (which adds voice cloning support), Hermes can now consume video input and produce spoken audio output within the same run.
That said, these features depend on external models and APIs. If you are running Hermes locally on a small open-source model, neither will work — the local models do not have multimodal inference capabilities. Worth being clear-eyed about that before planning workflows around them.
Where this connects to NordSys
Teams that want Hermes to do serious work — running overnight, handling sensitive data, switching providers as costs change — need the infrastructure underneath it to be solid. The provider architecture and security hardening in v0.13 bring it meaningfully closer to something you can run in production with confidence. We help businesses design and set up agent workflows that are reliable day-to-day, not just impressive in a demo. Our AI Agents service covers the full setup, from model selection to production configuration.
See our AI Agents service →