I keep a folder full of half-finished personal scripts. Snippets I started for one project, never quite generalized, never quite deleted. Every time I open the folder I think about cleaning it. Then I do not.
The thing that finally broke the loop, several years ago, was a backup tool that started flagging files I had not opened in a year. It did not delete them. It just put them in front of me with a date next to them. That was enough. Most went; a few stayed. The folder shrank.
I think about that tool when I read the v0.12.0 release notes. On April 30, 2026 — seven days after v0.11.0 rebuilt the interface — v0.12.0 shipped with a feature that, as far as I have seen, is the first time a self-hosted AI agent has had it: a background curator that grades and prunes your own skill library while you sleep.
The release notes call this the Autonomous Curator. Almost everything else in v0.12.0 — Microsoft Teams, Tencent Yuanbao, Spotify, Google Meet, LM Studio first-class, ComfyUI v5 bundled — would be the headline of a normal release. In this one, they share the spotlight with an agent that runs once a week to clean up after itself.
The Autonomous Curator
The Autonomous Curator is a background agent that runs on the gateway's cron ticker, with a default cycle of seven days. It walks your skill library, grades each skill against a rubric, consolidates near-duplicates, prunes the dead weight, and writes a markdown report to a known path with its reasoning attached.
Two technical details from the release notes worth pulling out. First: the self-improvement loop in v0.12.0 properly inherits parent runtime configuration, so the curator runs with the same model, sandbox, and credentials as the agent that spawned it. No surprise drift, no separate billing path. Second: the grading is rubric-based, not vibes-based — the curator rates skills against an explicit set of criteria you can read and, if you disagree with them, modify.
You did not opt in to the curator. It is already running. Next Sunday, your skill library will be smaller and sharper. You will get a report explaining what it threw out and why. If it threw out something you wanted, you can pull it back; hermes curator status ranks skills by usage so you can sanity-check the curator's call against your own.
The reason this is worth flagging as a category shift, not just a feature, is that every AI tool I have run has had a skills problem. You install a skill to try it. You forget you installed it. Six months later, you have a skills/ directory with two hundred entries, half of which actively contradict each other, all of which contribute prompt tokens. The agent itself never says "you should clean this up." Until v0.12.
The platform wave that almost got buried under the curator
Two new chat platforms shipped in v0.12.0. Microsoft Teams as the 19th messaging platform, shipping as a plugin — the full version, with end-to-end Graph auth + webhook listener + pipeline runtime + outbound delivery, lands in v0.14.0 a few weeks later. And Tencent Yuanbao (元宝) as the 18th, with native text and media delivery.
The pattern from v0.9.0 of "real consumer platforms" (iMessage, WeChat) is repeating, but with a twist: v0.12 leans toward enterprise chat. The Slack-and-Discord era of AI bots is well and truly over.
A flood of provider work
Model side, the headline is LM Studio getting promoted from "tested community path" to a first-class provider. If you have ever wanted to point Hermes at a local model and never look at a curl one-liner again, v0.12 is the release where that path finally became boring — in the good way.
Joining LM Studio: GMI Cloud, Azure AI Foundry, MiniMax OAuth, and Tencent Tokenhub. Four new providers in a release that everyone is going to remember as "the curator release" instead. Read the release notes carefully and you realize v0.12.0 is doing the work of three normal releases.
Spotify, Google Meet, and the integration tools
Two native integration tools showed up in v0.12.0 that deserve naming on their own. Spotify with PKCE OAuth — the agent can now skip songs, build a playlist, query what is playing on which device, all from a chat. Google Meet as a plugin that joins calls and pulls transcripts.
These are not headline features for everyone. They are headline features for the specific user who has been wondering whether "AI agent" could actually mean "controls things in the apps I use every day." For that user, v0.12.0 is the release that finally answers yes.
Skills got their own delivery channel
ComfyUI v5 and TouchDesigner-MCP graduated from optional to bundled-by-default. Seven new skills landed alongside: Humanizer (removes the AI-isms from output), claude-design, design-md, airtable, pretext, spike, and sketch.
The change that quietly opens a much bigger door is one line in the release notes: you can now install a skill directly from an HTTP(S) URL. There is no central registry required; if you can host a skill manifest, you can ship a skill. Combine that with the curator pruning your skills library every week, and you have the makings of a real ecosystem — one where people share skills the way they share gists, and the agent itself keeps the local library healthy.
A non-interactive mode, and the rest of the UX list
hermes -z <prompt> is the new non-interactive one-shot mode — fire a prompt, get an answer, exit. The release notes do not editorialize about this, but it is the feature that makes Hermes a real candidate for shell pipelines: echo "summarize this log" | hermes -z.
hermes update --check runs a preflight check before pulling. /reload-skills reloads the skill library mid-session. /busy lets you mark the agent as occupied, with a steering option attached. Visible TUI cold start shrank by roughly 57% thanks to lazy initialization and import-graph cleanup — small enough not to make the highlight reel, large enough that you notice on the first run after upgrade.
Breaking changes
Two to flag. Secret redaction now defaults to off (previously on) to prevent payload corruption — the prior default was occasionally eating valid tokens that pattern-matched the redactor. v0.13.0 will flip it back after the redactor itself gets reworked. And the /provider and /plan commands were dropped; their work was absorbed into /model and /steer respectively.
---
I keep going back to the curator. The whole "self-improving" framing around Hermes has been in the README since day one, and most of what it usually means in practice is: the agent gets better at writing skills. v0.12.0 introduced the other half of self-improvement. The agent is now better at removing skills, too.
This is a smaller thing to engineer than it sounds and a larger thing in shape. Every other release in this stretch made the agent able to do more. v0.12.0 made the agent able to do less, on its own initiative. The first kind of release expands the surface area you have to manage. The second kind shrinks it. The cadence of one followed by the other is, I think, the actual reason this project does not feel like it is collapsing under its own weight.