Hermes Agent 0.12.0 adds autonomous skill maintenance

Hermes Agent v0.12.0, released April 30, 2026, adds a background Curator that can maintain a user’s skill library without manual cleanup. The project says the new release also expands provider support, adds more messaging and voice integrations, and cuts visible TUI cold start by about 57%.

According to the release notes, the Curator runs as a background agent on the gateway’s cron ticker, with a 7-day default cycle. On each run, it grades skills, consolidates related ones, prunes dead entries, and writes reports to logs/curator/run.json and REPORT.md.

The project says archived skills are classified as consolidated or pruned using a model plus heuristics. It also says defense-in-depth gates protect bundled and hub skills from mutation, and that hermes curator status now ranks skills by usage. The Curator is managed under auxiliary.curator, with the model selected in hermes model or from the dashboard.

Hermes also says it substantially upgraded the self-improvement loop that reviews memories and skills after each turn. That background review is now class-first, meaning it uses a rubric instead of free-form judgment, and it is biased toward updating the skill the agent just loaded.

The release notes say the review fork now handles references/ and templates/ sub-files, inherits the parent runtime’s provider, model, and credentials, and is restricted to memory and skills toolsets. Hermes says memory providers now shut down cleanly, and prior-turn tool messages are excluded from the summary so the review sees a cleaner context.

On the integration side, the release bundles or promotes several tools. ComfyUI v5 moved from optional to built-in by default, and TouchDesigner-MCP is now bundled by default with expanded support for GLSL, post-FX, audio, geometry, and new reference docs.

Hermes also added native Spotify tools, a Google Meet plugin for joining calls, transcription, speaking, and follow-up, and a first-class LM Studio provider. The project says it added four more inference providers as well: GMI Cloud, Azure AI Foundry, MiniMax OAuth, and Tencent Tokenhub.

Gateway support is also growing. Hermes says Microsoft Teams is the first messaging platform shipped as a plugin, while Tencent Yuanbao becomes the 18th native messaging platform. The gateway now acts as a plugin host for drop-in messaging adapters.

The release includes a number of user-facing changes outside the new Curator. Hermes says hermes -z now supports one-shot non-interactive mode, hermes update --check adds a preflight check, and the models dashboard can show per-model analytics and switch main and auxiliary models in the browser.

Other changes include native multimodal image routing based on a model’s vision capability, broader multi-image sending across Telegram, Discord, Slack, Mattermost, Email, and Signal, pluggable TTS providers with Piper as a local option, and Vercel Sandbox as a terminal backend.

Hermes says secret redaction is now off by default to avoid patch corruption from secret-shaped substrings in tool output. The project also says visible TUI cold start was reduced by about 57% through lazy initialization, lazy imports, cached config loading, and other startup optimizations.

The release notes list 1,096 commits, 550 merged pull requests, 1,270 files changed, 217,776 insertions, and 213 community contributors since v0.11.0. Hermes Agent v0.12.0 is described by the project as the Curator release, with autonomous skill maintenance now built into the system.