Hermes v0.12.0 Features

Hermes Agent v0.12.0 ("The Curator release", NousResearch tag 2026.4.30) shipped 2026-04-30. 1,096 commits · 550 merged PRs · 213 community contributors since v0.11.0. Full upstream release notes: hermes-agent/RELEASE_v0.12.0.

This page captures the features that matter for AstroJOE / JOE deployments on the OPTX agent mesh — edge-node primary + containerized secondary + operator workstation. Where a feature is a thin wrapper over upstream behavior, the upstream PR is linked.

Autonomous Curator

Hermes Agent now maintains itself.

The hermes curator runs as a background agent on the gateway's cron ticker (7-day cycle by default). It grades the skill library, consolidates related skills, prunes dead ones, and writes per-run reports. This complements the AstroJOE self-maintenance protocol introduced in Hermes v0.8.0.

For OPTX: check ~/.hermes/logs/curator/REPORT.md after a Curator cycle. Pin skills you don't want curated (bundled and hub skills are already protected).

Self-improvement loop — substantially upgraded

The background review fork (the core of Hermes' self-improvement: after each turn it decides what memories/skills to save or update) got a series of fixes that make it more reliable on edge nodes.

• Class-first prompt — rubric-based grading rather than free-form (#16026)
• Active-update bias — prefers the skill the agent just loaded; handles references//templates/ sub-files (#17213)
• Fork inherits parent runtime — provider, model, and credentials propagate to the review fork (#16099)
• Scoped toolsets — review fork restricted to memory + skills (no shell, no web) (#16569)
• Clean shutdown — background review memory providers exit properly (#16204)

Bundled MCP servers — ComfyUI + TouchDesigner

Two creative-coding MCP servers moved from optional to bundled by default. AstroJOE across the agent mesh now ships with these tools without manual installation.

ComfyUI v5

Image generation via official CLI + REST + hardware-gated local install. (#17610, #17631, #17734)

TouchDesigner-MCP

Creative-coding bridge with GLSL, post-FX, audio, geometry, and 9 new reference docs. (#16753, #16624, #16768)

Inference providers — four new

Operational redundancy for the agent mesh. Configure via hermes model or the dashboard's new Models tab.

The remote model catalog manifest (#16033) means OpenRouter and Nous Portal model lists update without a Hermes release.

Messaging platforms

The gateway is now a plugin host. Drop-in messaging adapters live outside the core, and Microsoft Teams is the first plugin-shipped platform.

For the existing Matrix transport AstroJOE uses, see Optional Transport: Matrix Gateway.

Spotify + Google Meet — native integrations

Both ship as bundled plugins/skills, not bolt-ons.

• Spotify — 7 tools (play, search, queue, playlists, devices) behind PKCE OAuth, interactive setup wizard, surfaced in hermes tools, cron usage documented (#15121, #15130, #15154, #15180)
• Google Meet — join calls, transcribe, speak, follow up. Realtime OpenAI transport + Node bot server, full pipeline bundled as a plugin (#16364)

CLI quality-of-life

• hermes -z <prompt> one-shot mode — non-interactive runs with --model / --provider / HERMES_INFERENCE_MODEL. Useful for cron and scripted agent calls. (#15702)
• hermes update --check — preflight before applying updates, plus opt-in pre-update HERMES_HOME backup. (#15704, #15841)
• Models dashboard tab — rich per-model analytics, switch main + auxiliary models from the dashboard. (#17745, #17802)

TUI

• ~57% cold-start reduction via lazy agent init, lazy provider imports (OpenAI / Anthropic / Firecrawl), mtime-cached load_config(), memoized tool definitions, precompiled dangerous-command patterns (#17190, #17046, #17041)
• LaTeX rendering in transcript (#17175)
• /reload .env hot-reload (#17286)
• /mouse toggle — kills ConPTY's phantom mouse injection on Windows (#15488)
• Auto-resume of last session — opt-in (#17150)

Other notable upgrades

• Pluggable TTS provider registry — tts.providers.<name> registry with Piper as a native local TTS provider. Closes long-standing #8508. (#17843, #17885)
• Vercel Sandbox backend — execute_code/terminal can now run in a Vercel Sandbox (#17445)
• Native multimodal image routing — images route based on the model's actual vision capability rather than provider defaults (#16506)
• Gateway media parity — native multi-image sending across Telegram, Discord, Slack, Mattermost, Email, Signal; centralized audio routing with FLAC support (#17909)
• Configurable prompt cache TTL — prompt_caching.cache_ttl, default 5m, opt-in 1h for bursty sessions that keep cache warm (#15065)
• Secret redaction off by default — prevents long-standing patch-corruption incidents from fake secret-shaped substrings. Opt in via redaction.enabled: true. (#16794)
• Humanizer skill — strips AI-isms from text, ported from earlier branches (#16787)
• Bundled Langfuse observability plugin + hermes-achievements plugin that scans full session history (#16917, #17754)

---

Upgrading on the OPTX mesh

The operator workstation tracks the upstream Hermes Agent directly from a local clone — git pull && pip install -e . brings it to v0.12.0. The edge-node primary and containerized secondary are Docker-shipped; rebuild against the v0.12.0 image once the upstream release artifact is published. Until then, the deployed containers may still report an older Hermes — the operator workstation is the canonical source of truth for "current Hermes on this mesh."

For deeper detail on any item above, follow the linked upstream PR or read the full v0.12.0 release notes.