Hermes Agent v0.15.0 adds skill bundles and makes session search 750x faster

• NousResearch's release post announced Hermes Agent v0.15.0 as a large platform update, while Teknium's changelog summary said the release pulled in 747 PRs from 321 contributors.
• According to Teknium's summary, the headline speedups are 50 percent faster load times and session search that is 750x faster.
• NousResearch's MCP Catalog demo and the linked MCP Catalog docs show Hermes moving toward one-click installation of approved MCP servers, alongside the new skill bundles called out by Teknium's summary.
• Model coverage widened again: NousResearch's Krea announcement added Krea 2 as a built-in image provider, while Teknium's Opus 4.8 post and NousResearch's Qwen 3.7 Max post expanded the model list.
• The rollout was not perfectly clean. altryne's patch writeup documented an OpenAI Codex OAuth failure mode, Teknium's fix notice said an update resolved it, and Teknium's 0.15.1 post shipped a follow-up hotfix for dashboard and Kanban bugs.

The useful bits are all in the product surfaces. You can read the full release notes, browse the MCP Catalog docs, and watch the Atomic Bot app turn a self-hosted Hermes instance into an iPhone control surface. There is also a small rollout story hiding in the replies: OpenAI OAuth broke badly enough for users to post local patches before the official fix landed.

Hermes is getting more opinionated about packaging. Teknium's summary put skill bundles next to the new MCP Catalog, and NousResearch's MCP Catalog demo showed the catalog as a built-in flow rather than a docs-only feature.

The docs behind NousResearch's MCP Catalog link describe the catalog as one-click install for Nous-approved MCPs. That matters because it shifts setup from copying server configs by hand to picking from a curated list in product.

The release notes and tweets point to two separate packaging layers:

• Skill bundles, which Teknium's summary lists as a new release feature for grouping higher-level agent capabilities.
• MCP Catalog, which the official docs link frames as preconfigured MCP installs from trusted sources.
• Optional orchestration skills were already showing up just before the release, as rbren_dev's repost of the OpenHands skill highlighted.

The fastest concrete claims in the release are the easiest to remember: load times cut in half, session search sped up by 750x, and the old "godfile scripts" removed, according to Teknium's summary.

The same post groups performance work with security work, which gives a useful sense of what this release actually was: not one flagship feature, but a cleanup pass across the whole harness. The security list in Teknium's summary names three additions:

• Bitwarden native integration
• Brainworm prompt injection defense
• Automatic supply chain defense

Those are unusually operational features for an agent release. They sit closer to deployment hygiene than model capability, and they explain why this changelog reads more like infrastructure work than a simple feature drop.

Hermes keeps adding models faster than it adds messaging around them. NousResearch's Krea announcement added Krea 2 as a built-in image generation provider, while Teknium's Opus 4.8 post and NousResearch's Qwen 3.7 Max post extended the text model roster.

The release summary in Teknium's changelog thread bundles that expansion into a single list: Krea 2, Opus 4.8, Qwen 3.7, deep xAI integrations, and NFTY added to gateway channels. Teknium's provider reply also said Opus 4.8 was available through Nous Portal, direct Anthropic API access, and OpenRouter.

That leaves Hermes looking less like a single-model agent shell and more like a switching layer over multiple vendors and modalities. rohanpaul_ai's Grid demo pushed that logic even further by wiring Hermes into a market-routed provider layer instead of a fixed vendor endpoint.

The roughest part of the release window was OpenAI OAuth. bradmillscan's complaint thread described repeated gateway failures right after setup, and Teknium's reply blamed OpenAI rather than Hermes itself.

What made this more than ordinary launch-day grumbling is that altryne's patch writeup included a concrete diagnosis: Hermes could receive valid streamed output items from Codex, but the final response.completed frame could arrive without response.output, which then crashed openai-python 2.24.0 during final parsing. A few hours later, Teknium's fix notice said OpenAI had changed the spec and users needed to run hermes update to pick up the fix.

The cleanup continued after v0.15.0 itself. Teknium's 0.15.1 post said the hotfix existed partly for platforms that only work with tagged releases, naming Hostinger, and patched a dashboard load loop plus more Kanban adjustments.

One of the more interesting ecosystem moves around Hermes did not come from Nous at all. testingcatalog's Atomic Bot post showed an iOS app for controlling a self-hosted Hermes agent running continuously on a VPS, with remote access routed through Tailscale, Cloudflare, or ngrok.

testingcatalog's Atomic Bot link post says the agent still runs on your own server and the app is only the control surface. That split is a useful detail because it keeps the self-hosted posture intact while making Hermes look much closer to a consumer mobile product than most local-first agent stacks do today.