Hermes-Claude auth runs entirely through the official Claude Code CLI: you log in once with claude auth login, and the plugin never opens, copies, refreshes or prints those credential files — it simply drives the logged-in CLI, which also means every Hermes turn inherits, and bills against, whatever account claude is signed into. That one sentence is the entire security model. The rest of this page unpacks each clause of it, because when the account on the line is yours, the details deserve more than a shrug.
📺 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
Two sources feed everything below: the plugin's pinned README (version 0.3.0, dated 22 September 2026), which is the authoritative document on how auth actually works, and Julian Goldie's walkthrough video from the same day, which shows the login flow end to end. If you want the broader picture of what the plugin does, start with the Hermes-Claude overview — this page stays narrow: login, credentials and account safety.
Hermes-Claude Auth: Who Actually Holds the Keys
The README's answer is unambiguous: authentication belongs to the official Claude Code CLI, and to nothing else. The plugin holds no credentials of its own. It does not ask for a Hermes API key, and it never sends one. Where your Claude credential files are concerned, the README commits to four specific negatives — the plugin never:
- opens them,
- copies them,
- refreshes them, or
- prints them.
The plugin's whole job is to drive a CLI that is already logged in. Your tokens stay where the official tooling put them, managed by the official tooling.
The detail worth underlining is what happens when your environment disagrees. Plenty of tools quietly fall back to a direct HTTP path with an API key when their preferred auth route looks unavailable. The README says the client path does the opposite: there is no silent fallback. Before spawning the CLI, it actively rejects an inherited API key, a custom Anthropic endpoint, and cloud-backend overrides. The error names the conflicting environment variables without printing their values — so nothing sensitive lands in your scrollback — and the fix is exactly what you would want: remove those overrides when you select OAuth. You cannot accidentally bill an API key you forgot was exported.
The Login Flow, Step by Step
Julian's video walks the login live, and it is three moves:
- Install the official CLI: npm install -g @anthropic-ai/claude-code.
- Run claude auth login. Your browser opens, you approve the sign-in, and you paste the code back into the terminal.
- Done. The plugin finds the logged-in CLI by itself — there is no separate Hermes login.
What makes the auth experience feel engineered rather than bolted on is where the plugin checks its own setup: at every seam. Per the README, the checks run at plugin load, inside hermes model, in the pickers, and before every single request. Each failure refuses cleanly with the install hint instead of dumping a traceback on you. And on a terminal without a TTY, the login step does not try to guess its way through — it prints the instruction and stops.
Your everyday sanity check is claude auth status: a local read that comes back in roughly 0.3 seconds and confirms your credentials are valid, along with the plan they belong to — Claude Pro, in Julian's demo. No network round-trip, no side effects. The full install sequence, screen by screen, is in the subscription setup guide.
And if you would rather set this up alongside people who have already been through it, the AI Profit Boardroom is where Julian's community — 3,000+ members — trades working configs and flags the sharp edges before you hit them.
📺 Watch: Claude + Hermes Agent OS is WILD!
Whose Account Gets Billed
The README does not bury the billing question in a footnote; it gives it a Disclosure section of its own. Condensed to its substance:
Every turn is billed against your Claude subscription's Agent SDK allowance, and the provider inherits whatever account the CLI is logged into. Subscription entitlement and extra-usage settings belong to the account, not the plugin. Reported costs are native list-price equivalents, not verified subscription charges.
Three practical consequences for your account. First, identity: Hermes bills whoever is signed in, so on a shared or work machine, confirm before you run. Second, overage: extra-usage settings live in your Claude account, so if you do not want overage billing, disable extra usage there — no plugin setting can do it for you. Third, honesty about the numbers: the costs Hermes reports are list-price equivalents, an activity meter rather than an invoice, and the README says so plainly instead of letting you mistake them for verified charges. If you are weighing which models earn their cost per task, Goldie Bench is the companion reading.
Unsure whether subscription-backed auth or plain API billing fits your workload better? That is a fifteen-minute conversation — book a free strategy session and map it against what you actually run.
Model Discovery Without an Upstream Call
Here is a piece of credential hygiene most people miss. When you select the provider, the plugin needs to know which models your account can use — and it gets the answer without touching Anthropic. It asks the Claude CLI itself: the initialize handshake returns your account's own model picker, with zero upstream requests, which the README says was verified through the admission relay. Rows the CLI marks as drawing usage credits are labelled as exactly that, so nothing about cost hides inside the picker. And if the handshake fails, the plugin falls back to a pinned five-model catalog rather than erroring out or phoning home.
📺 Watch: Claude + Hermes Agent Just Changed SEO Forever!
Multi-Account and Non-PATH Setups
The README also covers the awkward auth situations — the ones you hit when one machine serves two lives.
- Separate accounts. Keep a second, CLI-managed auth directory: log in with CLAUDE_CONFIG_DIR set, then export CLAUDE_SUBSCRIPTION_DIRECTSDK_CONFIG_DIR to point the plugin at that directory. Your personal and work credentials never share a folder.
- CLI outside PATH. If the executable is not on your PATH, point CLAUDE_SUBSCRIPTION_DIRECTSDK_COMMAND at it directly.
- Per-profile installs. The plugin can be installed per profile under the profiles directory, so different Hermes profiles can carry different setups.
Headers and the Relay, in One Paragraph
Transport hygiene, per the README: requests move through a request-scoped loopback relay with a random per-request route, and native authorization headers pass through memory directly to the upstream — they are not logged and not persisted anywhere. Cancel a run and the plugin tears down the connection and the whole process tree with it. If you want the full mechanics of how the CLI gets driven, the transport deep-dive covers it end to end.
Before You Connect Your Account
This is the section to read twice, because it covers the one question the documentation cannot answer for you.
Here is what the docs do establish. The plugin ships in the official tier. It drives an unmodified official CLI. It wears an explicit Experimental label. It discloses, in plain language, whose account pays. And its credential handling — never opening, copying, refreshing or printing your files — is genuinely careful.
Here is what the docs do not establish: that Anthropic has blessed subscription use inside third-party harnesses. That is a terms-of-service question between you and Anthropic, and no plugin README — however careful — can settle it. Community members under Julian's video publicly asked exactly this, and his own reply is the right instruction: check Anthropic's current Terms of Service and your plan documentation before connecting your account.
Nobody can honestly tell you that accounts get banned for this, and nobody can honestly tell you it is cleared — neither claim appears in anything either source published. The current-terms check is your step, and it costs you five minutes.
Before your first billed turn, run this list:
- Run claude auth status and confirm which account you are about to bill — not just that some account is logged in.
- Juggling accounts? Give this setup a dedicated config directory via CLAUDE_CONFIG_DIR so a work login can never leak into personal experiments.
- Do not want overage billing? Disable extra usage inside your Claude account settings first.
- Treat the Experimental label as meaning exactly that — expect change, and re-read the README when the version moves.
Hermes is one piece of a larger stack, and the account you connect here powers the rest of it. The Agent OS guide maps the whole system, and Hermes inside Agent OS shows where this plugin sits in it.
Hermes-Claude Auth FAQ
Does the plugin see my Claude password or tokens?
No. Per the README, authentication belongs entirely to the official CLI. The plugin never opens, copies, refreshes or prints credential files, no Hermes API key is required or sent, and login itself happens in your browser through claude auth login. The plugin only drives the CLI that results — and it rejects API-key and endpoint overrides rather than silently falling back to them.
Which account does Hermes bill?
Whichever account the Claude CLI is signed into on that machine — every turn draws on that subscription's Agent SDK allowance. Entitlement and extra-usage settings live in the account itself, so the account owner controls overage. Run claude auth status before your first session; it takes a third of a second and names the plan and credentials you are about to use.
Is this allowed by Anthropic?
The docs do not settle it, and this page will not pretend they do. The README establishes official-tier status, an unmodified CLI and an explicit billing disclosure — but not that Anthropic endorses subscription use inside third-party harnesses. Follow Julian's own advice from his comments thread: check the current Terms of Service and your plan documentation before you connect, and make the call with the terms in front of you.
Get the Setup Right the First Time
Auth is the one part of a Hermes setup you do not want to learn by trial and error, because the account at stake is yours. Two shortcuts. Join the AI Profit Boardroom — 3,000+ members running this exact stack, $69/mo locked in while the price holds (normally $110). Or bring your machine's specifics to a free strategy session and leave with the login, account and billing questions answered for your setup, not a generic one.
