What's New in Hermes Agent: Kanban, Curator, Grok OAuth, and a Dramatically Lighter Install

The pace of Hermes Agent development over the past month has been extraordinary. Four numbered releases — v0.11 (Interface), v0.12 (Curator), v0.13 (Tenacity), and v0.14 (Foundation) — landed between late April and mid-May, totaling over 3,900 commits, 2,500 merged PRs, and more than 500 issues closed.

Here are the features you'll actually feel.

Multi-Agent Kanban: Delegate to an AI Team

v0.13 shipped one of the most requested capabilities: a durable multi-agent Kanban board.

Spin up a board, drop tasks on it, and let multiple Hermes workers pick them up, hand off, and close them out
Heartbeats, zombie detection, and retry budgets keep the team honest — if a worker goes silent, another claims the task
Hallucination gates and auto-block on incomplete exit prevent garbage from landing in your "done" column
One install, many kanbans — run separate boards for different projects

Combined with /goal (v0.13), which locks the agent onto a target across turns via the Ralph loop, Hermes is now far less likely to drift off-task mid-session.

Autonomous Curator: Your Agent Maintains Itself

v0.12 introduced hermes curator — a background agent that runs on the gateway's cron ticker and maintains your skill library on its own schedule:

Grades skills by quality and usage
Consolidates related skills into fewer, better ones
Prunes dead skills that haven't been used
Writes per-run reports so you can review what changed
Defense-in-depth gates protect bundled and hub skills from mutation
hermes curator status ranks skills by usage — most-used to least-used

The self-improvement loop (the core mechanism Hermes uses to decide what to remember after each turn) also got a major upgrade in v0.12: rubric-based scoring replaces free-form evaluation, the fork prefers the skill the agent just loaded, and live runtime credentials propagate correctly so the loop has access to the same tools the agent does.

xAI Grok via SuperGrok OAuth — 1M Context Window

v0.14 brought Grok into Hermes as a first-class provider. If you pay for SuperGrok, you can now use Grok inside Hermes by signing in with your xAI account — no API key, no separate billing.

grok-4.3 gets a 1M token context window — drop whole codebases or research corpora into a single prompt
OAuth flow handles entitlement errors gracefully
SSH-to-tunnel docs for completing OAuth when signed into a remote box

OpenAI-Compatible Local Proxy: One Subscription, Every Tool

Also in v0.14: hermes proxy starts a local HTTP endpoint that speaks the OpenAI API but is backed by whichever OAuth provider you're signed into — Claude Pro, ChatGPT Pro, SuperGrok.

The practical upshot: Codex CLI, Aider, Cline, Continue, and any custom script that expects an OpenAI-compatible endpoint now works with your existing subscriptions. No API keys required.

Debloating Wave: Lighter, Faster Installs

v0.14 made Hermes dramatically lighter:

Heavy backends now lazy-install on first use — messaging adapter SDKs, image-gen SDKs, and TTS providers are no longer pulled down at install time unless you actually use them
pip install hermes-agent works from PyPI — a clean install no longer drags in every optional dependency
Tiered install falls back gracefully when a wheel rejects on your platform
Cold start shaved by ~19 seconds — on top of the ~57% TUI cold-start cut from v0.12
Browser CDP calls are 180x faster — automation feels instant instead of laggy
Cross-session 1-hour Claude prompt caching — repeated inference is cheaper and faster

Video, Voice, and Media Tools

video_analyze (v0.13) — native video understanding on Gemini and compatible multimodal models. The agent can now watch and describe what's in a video
video_generate (v0.14) — unified pluggable video generation pipeline
xAI Custom Voices (v0.13) — voice cloning TTS provider
ComfyUI v5 + TouchDesigner-MCP (v0.12) — moved from optional to built-in by default

New Inference Providers

Nine new inference paths landed across the four releases:

xAI Grok via SuperGrok OAuth (v0.14)
GPT-5.5 via Codex OAuth (v0.11) — live model discovery in the model picker
Native AWS Bedrock (v0.11) — Converse API on a new pluggable transport layer
NVIDIA NIM (v0.11)
Arcee AI (v0.11)
Google Gemini CLI OAuth (v0.11) + native AI Studio API routing
Vercel ai-gateway (v0.11) — with pricing and dynamic discovery
OpenRouter Pareto Code router (v0.14)
Providers as a pluggable surface (v0.13) — add custom providers without touching core

Developer Experience

New Ink-based TUI (v0.11) — full React/Ink rewrite of the interactive CLI with a Python JSON-RPC backend. Sticky composer, live streaming, status bar with per-turn stopwatch and git branch
LSP semantic diagnostics on every write (v0.14) — real-time code feedback in supported editors
computer_use for non-Anthropic providers (v0.14) — the cua-driver backend now works beyond Claude
Clickable URLs in any terminal (v0.14)
Native Windows beta (v0.14)
Zed ACP Registry via uvx (v0.14)
huggingface/skills as a trusted default tap (v0.14) — community skills discoverable out of the box

Reliability and Security

Checkpoints v2 (v0.13) — rewritten state persistence with real pruning, so long-running sessions don't bloat
Gateway auto-resume (v0.13) — sessions interrupted by restart pick up where they left off
Redaction ON by default (v0.13) — 8 P0 security fixes closed in one release
Telegram native button UI for clarify (v0.14) — the agent can now ask you clarifying questions with tappable buttons instead of text walls

Internationalization

Hermes v0.13 shipped static gateway and CLI translations for 7 locales: Chinese, Japanese, German, Spanish, French, Ukrainian, and Turkish. The docs site also gained a Chinese (zh-Hans) locale.

If you want to try the latest Hermes Agent without managing your own server, ClawSimple now offers one-click Hermes deploys with runtime switching between OpenClaw and Hermes from your Profile page.