Hermes Agent ships v0.14.0 with Grok subscriptions, Codex runtime, and Windows beta

• According to NousResearch's release notes screenshot, Hermes Agent v0.14.0 bundles Grok OAuth access, first-class X search, an opt-in Codex runtime, model-agnostic computer use, native video generation, Windows beta support, and new gateway channels in one release.
• NousResearch's X Premium+ announcement and NousResearch's SuperGrok post show Hermes now accepts both X Premium+ and SuperGrok subscriptions, which unlock Grok models plus X search without requiring a direct xAI API key.
• In WesRoth's Codex runtime screenshot, Hermes can hand OpenAI turns to the Codex CLI app-server so terminal commands, file edits, sandboxing, and MCP calls run inside Codex's runtime while Hermes stays the outer shell.
• the release notes screenshot says the release also adds Microsoft Teams, LINE, and Simplex Chat support, bringing Hermes to twenty-two messaging platforms, plus a native Windows path with cmd.exe and PowerShell instead of WSL.
• Early hands-on posts from niallohiggins' migration note and LLMpsycho's overnight jobs post frame Hermes less as a single model frontend and more like an agent harness with reusable profiles, slash commands, and installable skills packs.

You can read the GitHub release notes, the xAI integration post, the X search docs, and the Grok OAuth guide. The weirdly practical bit is that Hermes now spans both sides of the closed-model market, with Grok arriving through subscription OAuth while OpenAI gets a separate Codex app-server runtime. There is also a surprisingly broad surface area hidden in the patch notes, from Zed registry install support to an optional skills catalog that already includes MLOps-specific packs.

The biggest access change landed before the full 0.14.0 drop. NousResearch's X Premium+ post says Hermes can now use X Premium+ subscriptions, while its SuperGrok announcement added SuperGrok accounts a day earlier.

That path matters because the auth flow carries more than just model access. According to Teknium's summary, Grok login in Hermes unlocks Grok, X search, image generation, video generation, and voice, and altryne's thread adds one limitation: this is not full X API access, so features like bookmarks still need a separate API key.

The X search tool is a distinct product surface, not just a logged-in X client. testingcatalog's doc screenshot says x_search is backed by xAI's Responses API and returns synthesized results with citations, while altryne's example list frames it for public-X research tasks like reaction sweeps, thread hunting, and launch summaries.

The OpenAI side of the release is more architectural. In WesRoth's screenshot, Hermes can optionally route openai/* and openai-codex/* turns to the Codex CLI app-server instead of using Hermes' own tool loop.

That split leaves Hermes acting as the wrapper around sessions, slash commands, gateway routing, memory, and skill review, while Codex handles the inner execution path for terminal commands, file edits, sandboxing, and MCP tools. WesRoth's later repost repeats the same runtime description and makes clear the flag is opt-in.

This release ends up looking like a dual-access bet. a Grok provider selection screenshot reposted by Teknium shows XAI Grok OAuth sitting alongside OpenAI Codex and a long provider list, which is Christmas come early for agent framework tinkerers who want one shell wrapped around several tool runtimes.

Much of 0.14.0 is pure surface-area expansion. The release notes screenshot lists three new gateway channels, Microsoft Teams, LINE, and Simplex Chat, and says Hermes now spans twenty-two messaging platforms.

The same notes pack in several deployment changes:

• Grok via SuperGrok, with a one million token context window for grok-4.3, per NousResearch's release notes screenshot
• Native Windows support in early beta, with cmd.exe and PowerShell instead of WSL, per the same notes
• One-click install from Zed through the Agent Client Protocol registry and a new uvx install path, per the screenshot
• Slash /subgoal and live /handoff, which let users add success criteria mid-run and move sessions between models or profiles without losing context, per the screenshot

The provider-picker screenshot in Teknium's repost also hints at the project's ambition creep. Hermes is no longer presenting itself as a narrow shell for one model family, it is shipping as a routing layer across local, API, OAuth, and gateway-backed providers.

The patch notes close with the kind of details operators actually bookmark. NousResearch's release notes screenshot says Hermes added nine optional skills, closed twelve security issues across approval gates, sudo brute force, and SSRF, cut cold start by 19 seconds, and made browser tool calls 180 times faster.

The skills system already goes well past toy examples. In mervenoyann's catalog screenshot, the optional skills list includes huggingface-accelerate, axolotl, chroma, faiss, Flash Attention optimization, and Hugging Face tokenizers, which makes the catalog read more like a practical plugin layer than a prompt library.

Security is showing up in community usage too. dangtony98's post says one setup runs Hermes without giving it any credentials at all, while a later repost from the same account amplifies concern that Hermes can read raw API keys. That tension, richer automation versus tighter secret handling, is becoming part of the product story around open agent shells.

The early user reports are mostly about operations, not benchmarks. niallohiggins' migration thread says moving from OpenClaw to Hermes was easier than expected and that coffee ordering, Amazon subscription management, financial skills, agent mail, Stripe Link CLI, and Bitwarden secrets all ported cleanly.

The same thread breaks the perceived differences into a short list:

• Approvals feel more explicit, according to niallohiggins' follow-up
• Cron and job management are clearer, according to the same follow-up
• Workflow triggers are easier to reason about, according to that post
• Telegram plus slash commands make mobile parallel workflows practical, according to niallohiggins' next reply

Other users are already talking about Hermes as an installable environment. LLMpsycho's post compares the pack install flow to oh-my-zsh for agents and says it reduced repeated setup work for overnight jobs, while altryne's ThursdAI note says all ThursdAI cohosts had moved to Hermes and Codex.

The release was paired with a broader push onto new surfaces outside the CLI itself. NousResearch's NVIDIA post says the team worked with NVIDIA so Hermes runs on DGX Spark, and an Ollama repost of NVIDIA AI PC points to a playbook for running Hermes fully locally through Ollama on that hardware.

That sits next to two quieter ecosystem pieces from the same weekend. Teknium's Atlas post calls Hermes Atlas a useful resource, and his follow-up link points users to the main Hermes site, which now acts as the docs, install, and integration hub for the expanding stack.

Taken together, 0.14.0 looks less like a patch release than a foundation pass. The notes in NousResearch's screenshot count 633 PRs from 215 contributors, and that number makes sense once you add up the real payload: new auth paths, new runtimes, new gateways, a Windows port, faster tools, and a growing skills layer.