Hermes Claude is a live pairing again — after Claude stopped being usable inside Hermes, NousResearch shipped an official plugin on 20 September 2026 that runs Claude through your existing Claude Code subscription, and it slots straight into Hermes' tools, memory and workflows.
📺 Watch: New Hermes + Claude Plugin Update is AWESOME!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
That sentence hides a lot of history. For a while, the honest answer to "can I run Claude inside Hermes?" was no. Now it is yes — officially, from Hermes' own maker, on the subscription you already pay for. This page is the hub: the timeline of what broke and what came back, the models and workflows you can run today, how the plugin works at a glance, the one caveat worth reading before you connect, and where each deep-dive lives when you want the full detail.
Hermes Claude: what broke and what came back
First, the break. Hermes users used to be able to run Claude, and then they could not. Julian Goldie covered the whole arc in his video "New Hermes + Claude Plugin Update is AWESOME!" (22 September 2026), and his summary of the shutdown era is blunt: "they actually stopped it." If you lived through that window you remember the shape of the problem — your preferred agent harness on one side, your preferred models on the other, and no supported bridge between them. Plenty of people quietly gave up on the combination.
The fix arrived on 20 September 2026 as an official catalog plugin. The id is claude-subscription-directsdk, it appears in the catalog as Claude Subscription DirectSDK (Experimental), and the publisher is NousResearch — the maker of Hermes itself, not a third-party shim vendor. The code was extracted from hermes-agent PR #105863, with core support landing in PR #117451, and the README credits the pull request authors @teknium1 and @unsupportedpastels for the work. In other words: this is the first-party answer to the question everyone kept asking.
Getting in is one line: hermes plugins install claude-subscription-directsdk. After that it is a sign-in and a model pick, and the setup deep-dive below walks through both step by step.
What you can run now: models, subscription, one door
With the plugin installed, Claude models show up as selectable Hermes providers on a Claude Pro or Max subscription — no separate API key, no per-token billing bolted on the side. Five models are on the menu today: Sonnet 5, Haiku 4.5, Opus 5, Opus 4.8 and Fable 5.1. Switching between them is a one-liner with hermes model, or a dropdown click if you drive Hermes through a chat interface. Pick the heavy model for hard reasoning, the fast one for volume work, and move between them mid-project.
Julian's gym analogy is the cleanest way to say why this matters. Running Claude on an API key while already paying for a Claude subscription was like paying a gym membership and then per-minute fees for the good machines. The plugin removes the second charge. One subscription, one door, every machine on the floor.
The half of the announcement that matters most is what Hermes keeps. Its agent loop, tools, approval prompts, memory and compaction all continue to work with Claude underneath — you are swapping the engine, not the car. If you want to understand what that memory layer actually does between sessions, the Hermes memory system breakdown covers it in depth.
Julian runs this in production on his own machine: Claude inside his Agent OS chat under a profile named Claude Opus, powering the same workflows he was already running — Hermes Oracle, Muse and Astros all now run with Claude underneath. Nothing about those workflows changed except the model behind them. To build the same foundation, start with the Agent OS guide, then the Hermes Agent OS walkthrough.
And if you are building on this stack, you do not have to figure it out alone. The AI Profit Boardroom is where 3,000+ members share working Hermes and Claude setups, Agent OS templates and the prompts behind them. Want a plan tailored to your business before you touch anything? Book a free strategy session and map it out one on one.
📺 Watch: Claude + Hermes Agent OS is WILD!
How the plugin works, in one section
Hub-level only here — the transport deep-dive owns the detail. The core design decision: the plugin does not reimplement Claude or ship its own client. It drives the unmodified, official Claude Code CLI already installed on your machine. Authentication stays where it always was: you run claude auth login once, the CLI holds your credentials, and the plugin never opens or copies them. Hermes talks to the CLI, and the CLI talks to Anthropic.
Between the two sits a request-scoped local relay with a deliberately narrow job: it admits exactly one upstream request per call, and nothing else rides along. On costs, know that the numbers Hermes reports are list-price equivalents — useful for comparing one run against another, but not verified subscription charges and not an invoice.
The README (v0.3.0) deserves credit for honesty. It wears the Experimental label openly and describes its own security review as exactly that — a review build, not a production-readiness claim. Plugin pages rarely undersell themselves; this one does, and that self-awareness cuts both ways: trust the design, respect the label.
One line on plan rules, because it decides which subscription you need: Fable 5.1 on a Pro plan draws usage credits under Anthropic's documented plan rule, while Max plans include Fable up to 50% of weekly limits. The exact numbers, and what they mean for daily driving, live in the setup deep-dive below.
📺 Watch: Claude + Hermes Agent Just Changed SEO Forever!
Before you connect
Two caveats, briefly, because pretending they do not exist helps nobody. First, Nous marks the plugin Experimental and ships an explicit billing disclosure: every turn bills your subscription's Agent SDK allowance, and the provider inherits whichever account is logged in on the machine. Second, the terms question. Community comments under Julian's video asked whether routing a subscription through a third-party harness fits Anthropic's current terms, and Julian's own reply was the right one: check Anthropic's current Terms of Service before connecting. That is the honest status — an official plugin, an Experimental label, and a question you should answer for yourself rather than take from a blog post in either direction. The auth deep-dive below covers the full picture, including exactly what the plugin can and cannot see.
Where each deep-dive lives
This page is the map. The territory is split across three siblings, all live now:
- Full setup — prerequisites (Hermes 0.21.4), the install command, sign-in, model profiles and the plan limits in detail: Hermes agent with your Claude Code subscription.
- Auth and credentials — how login actually works, where credentials live, what the plugin never touches, and the terms question in full: Hermes Claude auth, explained.
- The CLI transport — what driving the unmodified Claude Code CLI means in practice, the one-request relay, and how cost reporting works: the Hermes Claude CLI transport.
Read them in that order if you are new to the Hermes plus Claude stack. If you only read one, read the setup guide — it takes you from a bare Hermes install to a working Claude provider in one sitting.
Hermes Claude FAQ
Can Hermes use Claude again?
Yes. Since 20 September 2026, the official claude-subscription-directsdk plugin from NousResearch — Hermes' own maker — restores Claude as a first-class Hermes provider, running through your existing Claude Code subscription instead of a separate API key. It carries an Experimental label, so expect rough edges, but it is the supported path and it works today.
Which Claude models work in Hermes?
Five Claude models are selectable as Hermes providers right now: Sonnet 5, Haiku 4.5, Opus 5, Opus 4.8 and Fable 5.1, on a Claude Pro or Max subscription. Switching is a single hermes model command or a dropdown change. For how those models actually compare on real agent workloads, see Goldie Bench.
Is it free?
There is no new charge on top of your subscription, but it is not free compute either. Every turn draws from the Agent SDK allowance of the Claude plan you already pay for, so heavy Hermes sessions pull down the same weekly limits as the rest of your Claude use. The cost figures Hermes displays are list-price equivalents shown for comparison — not actual charges to your account.
Get Hermes and Claude running today
The door between Hermes and Claude is open again, and the people moving fastest through it are the ones comparing notes instead of debugging alone. Inside the AI Profit Boardroom you get the working setups, Agent OS templates and prompts from 3,000+ members, at $69/mo locked in — normally $110. Prefer to talk it through first? Grab a free strategy session and we will map your Hermes Claude rollout — subscription, models, workflows — before you spend a minute on trial and error.
