Hermes Agent on your Claude Pro or Max plan, through Claude Code
Nous's experimental plugin runs Hermes turns through your own Claude Code on Pro or Max. Setup on a server, extra_usage, Anthropic's terms, "Unknown provider".
Install Nous Research’s official plugin claude-subscription-directsdk. It is experimental, and it runs each Hermes turn through the official Claude Code CLI, signed in with your Claude Pro or Max plan, so the login stays with Claude Code. It needs Hermes 0.21.4 or newer and Claude Code installed and logged in. Install Claude Code, install the plugin, run claude auth login, then pick “Claude Subscription DirectSDK (Experimental)” in hermes model. On a server over SSH, the login gives you a URL to open in any browser and takes back the code it shows. Per the plugin’s docs, “every turn is billed against your Claude subscription’s Agent SDK allowance.” This is a different route from Hermes’s own Anthropic sign-in, which works only on Max with extra usage bought and bills everything as extra usage. Whether Anthropic accepts either route is Anthropic’s call: its terms let you sign in to the unmodified Claude Code with your own plan, and bar other developers from routing requests through plan credentials for their users.
npm install -g @anthropic-ai/claude-code
hermes plugins install claude-subscription-directsdk
claude auth login
hermes model # pick "Claude Subscription DirectSDK (Experimental)"CopiedWhat the plugin is
The plugin’s docs describe it as an “Experimental model provider that runs Hermes turns on a Claude Pro/Max subscription through the official Claude Code CLI”. Its README says it “drives the unmodified official Claude Code executable” and talks to it directly, without the Python Agent SDK package. The agent stays Hermes: “Hermes then applies its own hooks, approvals, tools and persistence.” The plugin keeps no credentials of its own and “never opens, copies, refreshes, or prints its credential files”; whatever account claude is logged into is the one Hermes uses.
Nous calls it “a review build, not a full-parity or production-readiness claim.” Version 0.4.0 reached catalog installs on November 19, 2026. The catalog install is pinned to a reviewed version; installing from NousResearch/hermes-official-plugins#claude-subscription-directsdk tracks the repository’s main branch instead. No plugins.enabled entry is needed. On a Hermes older than 0.21.4 it does not half-work: it “fails to load with a clear Failed to load user provider plugin warning”. Update Hermes first; the current stable release on November 16 was v0.23.2.
Each model needs a recent enough Claude Code. The docs list Opus 5.5 at 2.1.280 or newer, Sonnet 5.5 at 2.1.284, Haiku 5.5 at 2.1.293 and Fable 5.2 (the fable alias) at 2.1.321, and warn that “An older CLI answers 400 Claude Code … does not support this model, so run claude update first”. If claude update stops short, the “does not support this model” guide covers release channels.
Signing in on a server with no browser
The setup in hermes model checks claude auth status. On a terminal, a logged-out check starts claude auth login for you; “Without a TTY it prints the instruction and stops”, and you run claude auth login yourself. An SSH session is a terminal, so either way you end up in Anthropic’s own login. Per Anthropic’s authentication docs, when no browser opens, press c to copy the login URL and open it in a browser on any device. If that browser shows a code instead of redirecting back, paste it at the Paste code here if prompted prompt. The terminal then shows Login successful.
On Linux the login is stored in ~/.claude/.credentials.json, readable only by your user. The Hermes gateway has to run as that same user to use it. When the gateway runs under systemd, launchd, a container or cron, Claude Code without USER in its environment reports loggedIn: false even when it is logged in; the plugin now sets USER itself before starting Claude Code, and a logged-out error carries Claude Code’s own first lines of output. To use a different Claude login for a Hermes profile, the plugin’s claude_config_dir setting points it at another Claude Code configuration directory.
One thing outranks the stored login. “Claude Code prefers CLAUDE_CODE_OAUTH_TOKEN over the login in its config directory, so the token decides first,” in the plugin docs’ words, and Anthropic’s precedence list agrees. That variable holds the one-year token claude setup-token prints, which Anthropic documents for “CI pipelines, scripts, or other environments where interactive browser login isn’t available”; what it can and cannot do is in the setup-token guide. If Hermes runs as the wrong account, check whether that variable is set where the gateway runs. When Claude Code itself says it is logged out or its login expired, the “Not logged in” guide has each message and its fix.
What the usage draws on
The plugin’s own line is that “every turn is billed against your Claude subscription’s Agent SDK allowance”, and its README adds that “Subscription entitlement and extra-usage settings still belong to the account and native service.” Two details from its docs:
- Past the plan’s limit. The
extra_usagesetting decides “What a request does once Claude Code reports that this account has passed its plan limit and is drawing paid extra usage”.allow: “nothing changes.”notify, the default: “requests continue; one warning per limit window goes to the log”.ask: “requests are refused before Claude Code starts”, with how to allow extra usage until the reset.stop: “requests are refused until the plan limit resets”. So with the default, a Hermes left running keeps spending extra usage if your account has it turned on. - Fable on Pro. “on Pro (and Team standard seats) Fable 5 / 5.1 bill to pay-as-you-go usage credits from the first request.”
hermes config set plugins.entries.claude-subscription-directsdk-experimental.settings.extra_usage stopCopiedThat is the reverse of Hermes’s built-in Anthropic sign-in, which per Hermes’s provider docs “only works if you’re on a Claude Max plan and have purchased extra usage credits” and leaves the plan’s included allowance untouched. The provider docs do not mention the plugin. Hermes, Pi and OpenCode’s routes to a Claude plan are compared in your Claude subscription in Hermes, Pi or OpenCode.
What Anthropic’s pages say
Anthropic’s legal and compliance page calls subscription sign-in “designed to support ordinary use of Claude Code and other native Anthropic applications”. It says “Anthropic does not permit third-party developers to offer Claude.ai login into their own applications, or to route requests through Free, Pro, Max, or Ultra plan credentials on behalf of their users,” and that “sign-in to a Claude account must complete through Anthropic’s own flow.” It does not “prevent an end user from signing in to the unmodified Claude Code binary with their own Claude subscription,” and Anthropic may enforce “without prior notice.” The Agent SDK overview adds: “Unless previously approved, Anthropic does not allow third party developers to offer claude.ai login or rate limits for their products, including agents built on the Claude Agent SDK.”
On billing, Anthropic’s Agent SDK article said on June 15, 2026 that “Claude Agent SDK, claude -p, and third-party app usage still draw from your subscription limits,” without naming the apps. Its help article on logging in says Anthropic “may at its discretion allow paid subscribers who have enabled usage credits to use certain third-party tools”, and prohibits tools that “misrepresent their identity to Anthropic’s servers” or “attempt to route third-party traffic against subscription limits”.
Set beside those: the plugin runs the unmodified Claude Code binary, the login completes in Anthropic’s own flow, and the plugin never touches the stored credentials. Claude Code names its caller on every request, and the plugin names it hermes-agent. What it sends against your plan is the work of Hermes, an agent from another developer. Neither the plugin’s docs nor its README say anything about Anthropic’s terms. Which side of these lines the route falls on is for Anthropic to decide, and an Anthropic API key in Hermes stays the route Anthropic’s help article calls preferred for other tools.
“Unknown provider ‘claude-subscription-directsdk-experimental’”
Unknown provider 'claude-subscription-directsdk-experimental'. Check 'hermes model' for available providers, or run 'hermes doctor' to diagnose config issues.CopiedHermes is set to the plugin’s provider but could not find it at that moment. Three open issues in November 2026, all on Hermes v0.23.2 with plugin 0.3.5, show two causes:
- Several chats of one profile opening at once. In #156899 and #157976, a profile other than the default, served by
hermes serveto the desktop app, lost the provider when several lookups ran at the same time, often right after a restart. A Hermes collaborator named PR #157087 as the fix, measured at 10 failed turns in 30 before it and none after; on November 19 it was still open, so the fix is a Hermes update once a release carries it. One reporter found a gateway restart did not reliably clear the error. - A profile that never had the plugin. In #157647, a profile created through Hermes’s API inherited the main profile’s model choice, but the plugin was installed only in the main profile. The reporter’s workaround was to make the plugin present in the new profile’s
plugins/directory and restart the gateway.
In either case, hermes model in the affected profile shows whether Hermes sees the provider. Installing Hermes, running its gateway as a service and choosing a provider are in the Hermes install guide.