English

← Back to list

Hermes Agent v0.14.0: One Subscription Powers Every AI Tool

itadol5j

Cover

If we're naming the most restless open-source project in the AI world this May, Hermes Agent deserves a spot. Nous Research's v0.14.0 is what the team itself calls the "Foundation Release." One-line summary:Hermes now installs anywhere, ships the features you actually want, and stops cramming in stuff you'll never use.

How substantial is it? From v0.13.0 to v0.14.0: 808 commits, 633 PRs, 1.393 million lines changed, 545 issues closed (including 12 P0s and 50 P1s), 215 community contributors. That's essentially a full reforge of Hermes.

And the performance numbers are even juicier—browser calls 180x faster, 19 seconds shaved off startup, Claude prompt caching across sessions for 1 hour. Let's dig into each one.

pip install hermes-agent Finally Works

Installing Hermes used to mean either cloning the repo or running a shell install script. With v0.14.0, Hermes has landed on PyPI:

pip install hermes-agent
hermes

Done. The Ink TUI and shell launcher are all bundled in the wheel—full experience out of the box.

It looks small but means a lot—it means Hermes can now drop into any Python-ecosystem CI/CD pipeline, Dockerfile, or conda environment. No more "installing Hermes is like installing Linux" jokes.

xAI Grok Joins the Family, grok-4.3 Gets 1M Context

If you subscribe to SuperGrok (xAI's paid tier), you can now log in with Grok via OAuth directly—no API key, no separate billing.

Even more exciting:grok-4.3 gets pushed to a 1M-token context window in Hermes. What does that mean? The entire fs/ directory of the Linux kernel source, the whole Three-Body Problem trilogy in one volume, dozens of papers—all of it fits in a single feed.

For anyone doing large codebase analysis, long-document summarization, or research tasks, this is a cheat code. There's even thoughtful "SSH-to-tunnel" documentation for remote scenarios—run hermes on a remote server and still complete OAuth.

One Command Turns Every Subscription Into an OpenAI Endpoint

Illustration

This is the slickest feature of the release:hermes proxy.

Run that command and a local OpenAI-compatible http://localhost:port endpoint spins up. Behind it sits any Hermes provider you've OAuth'd into—Claude Pro, ChatGPT Pro, SuperGrok, all of them work.

And then what?

  • Codex CLI wants in? Point it there and it just works
  • Aider wants in? Point it there and it just works
  • Cline / Continue / your own scripts? Every one of them works

One subscription feeds every tool. No more opening separate API keys and topping up balances for each tool. This move strikes right at the heart of the OpenAI/Anthropic business model.

x_search: Twitter Search Becomes a Native Tool

Getting an agent to search Twitter used to mean installing a Skill, configuring an API, writing a wrapper. Now x_search it's a first-class citizen tool:

  • search timelines
  • find thread posts
  • pinpoint specific posts

OAuth login or API key, whichever suits you. For sentiment monitoring, trend tracking, and KOL analysis, this thing is a cheat code.

Microsoft Teams, Full-Stack Integration

This release wires up the entire Teams stack:

  • Microsoft Graph authentication
  • Webhook listener
  • Pipeline plugin runtime
  • Outbound message delivery

In practice: configure a bot once, and your agent can be @-mentioned to get work done in any Teams channel, DM, or group. A godsend for enterprise users—plenty of companies run on Teams internally.

Add in the newly added LINE and SimpleX Chat, and the total number of messaging platforms Hermes supports now reaches 22. WeChat, Feishu, WeCom, DingTalk, QQ, WhatsApp, Discord, Slack, Telegram, Signal, Email... basically everything you can think of.

Debloating: Installing Hermes No Longer Means Dragging Along What You Don't Use

Illustration

Older versions pip install hermes-agent pulled down every messaging adapter, every image generation SDK, every voice/TTS provider. The result was a Python version of the node_modules black hole.

v0.14.0 changes that:

  • Heavyweight backends (Slack, Matrix, Feishu, DingTalk adapters, the hindsight client, codex app-server, Pixverse, Camofox, image-gen SDKs, voice/TTS providers) are nowauto-installed on first use
  • [all] extras no longer duplicate what lazy-deps already covers
  • when a wheel fails to install, the installer falls back automatically by tier
  • every install runs asupply chain vulnerability scan

The result:

  • faster installs
  • smaller disk footprint
  • far fewer transitive dependency vulnerabilities

Three Waves of Performance Gains

Startup Is 19 Seconds Faster

hermes This command used to take several seconds before the prompt appeared—a pile of import overhead plus network calls. v0.14.0 basically reworked the startup path to deferred (load on demand),cutting cold start by about 19 seconds.

Browser Calls Are 180x Faster

When the agent uses the browser tool to inspect pages or run JavaScript, it used to spin up a fresh DevTools session every single time. Now it'sa single persistent Chrome connection that gets reused. Calls that took seconds now return in milliseconds. Page interactions feel visibly "buttery."

Claude Prompt Cache Lasts 1 Hour Across Sessions

When using Claude via Anthropic, OpenRouter, or Nous Portal, the prompt prefix (system prompt, Skills, memory) is cached across sessions for 1 hour./new Start a new session and the first response is both fast and cheap. Background memory recalls hit the cache too—no more paying full price on every turn.

Other Features Worth Mentioning

  • /handoff: you can now genuinely hand a session off to another agent in real time
  • Telegram / Discord native button UI: respond to clarify prompts by tapping buttons instead of typing
  • Discord channel history backfill: newly connected channels can pull in their chat history
  • LSP semantic diagnostics on every file write: code gets checked the moment it's written
  • A unified, pluggable video_generate: video generation is no longer locked to a single vendor
  • computer_use cua-driver backend: computer use finally runs on non-Anthropic models
  • Clickable URLs in the terminal: open links directly from any terminal
  • Zed ACP Registry integration via uvx: great news for Zed users
  • Native Windows beta: no WSL2 dependency
  • 9 new optional Skills
  • OpenRouter Pareto Code router: automatically picks the code model with the best value for money
  • huggingface/skills added to the default trusted sources

One command to migrate over from OpenClaw

Illustration

This is Hermes's dirtiest joke—and its deadliest weapon.

hermes claw migrate

It moves your entire OpenClaw setup—config, data, workspaces—straight into Hermes. Combined with hermes profile for separating work, personal, and test environments,switching costs are basically zero.

Plenty of people in the community have already gone the "dual-wield" route: OpenClaw handles connections and scheduling, Hermes handles the thinking and learning for complex tasks. One dual-wield user put it like this:"OpenClaw does the work, Hermes does the thinking."

The Fastest Way to Install Hermes

Users in China can use a mirror for faster downloads:

Linux / macOS / WSL2:

curl -fsSL https://res1.hermesagent.org.cn/install.sh | bash

Windows PowerShell:

irm https://res1.hermesagent.org.cn/install.ps1 | iex

Then:

hermes setup    # 配置模型（GLM、Kimi、MiniMax、Claude 都行）
hermes          # 启动 TUI
hermes gateway  # 启动消息网关

60 seconds from zero to chatting.

Why v0.14.0 Deserves Your Attention

Hermes Agent went open source in February this year, and in just over two months its GitHub has climbed to 71,000 stars, peaking at 6,400 stars in a single day. Growth like that is rare even by open-source standards.

v0.14.0 sends three core signals:

First, Hermes has evolved from "experimental project" into "engineering-grade product." The PyPI release, debloating, install-time vulnerability scanning, and native Windows support are all hard metrics aimed at real production environments.

Second, Hermes's differentiation is getting clearer—it's not fighting OpenClaw for market share, it's offering a more focused choice for users who want "AI that learns to do things." OpenClaw is like Android, giving you maximum freedom; Hermes is like Apple, turning experience into habit for you.

Third, that hermes proxy design is a genuine masterstroke—one Claude Pro / ChatGPT Pro / SuperGrok subscription feeds every OpenAI-compatible tool. That's not just a feature; it's cracking open a business model.

If you're an indie developer, a personal-workflow geek, or just curious about "a different AI agent philosophy," v0.14.0 is a great version to jump in with.

If you're an enterprise user or a stability-sensitive team, it's fine to wait and watch—but the Microsoft Teams integration signals that Hermes is actively moving toward enterprise scenarios too.

Whatever you choose, the battle of open-source AI agents has only just begun in 2026.