Skip to main content
Everything PR News
Technology

Hermes and the Claude Harness Detection Incident Explained

EPR Editorial Team7 min read
Share
hermes claude ai platform governance explained overview video game media

Hermes is a third-party Claude Code harness built by Nous Research. In April 2026, a bug in Anthropic's third-party harness detection silently moved Claude Code subscribers to pay-as-you-go billing when the string HERMES.md appeared in a Git commit message, and an Anthropic engineer publicly acknowledged it. The episode became a test case for AI platform governance and a map of how third-party harnesses relate to Claude.

By the EPR Editorial Team. Originally published May 27, 2026. Updated October 10, 2026. Incident facts are drawn from public reporting and Anthropic's public acknowledgment. Platform behavior changes quickly, so verify technical specifics against primary sources before acting on procurement or engineering decisions.

What is Hermes?

Hermes is a terminal-based harness built by Nous Research that wraps Anthropic's Claude Code CLI and runs it inside its own terminal, with its own skills library, credential routing, and multi-turn orchestration. Claude Code is Anthropic's product, and Hermes is a developer tool that uses it, the way Cursor or Aider use a model. It is not a competitor to Claude but a different way of consuming Claude, built for longer and more autonomous coding sessions than the native interface was designed for. The practical comparison is in Hermes vs Claude Code: When to Use Which, and the setup problems most users hit are in Hermes Auth Setup.

Tech PR for brands that ship5W helps technology companies win coverage, analyst attention and a place in AI answers.See 5W tech PR212.999.5585info@5wpr.com

What happened in the Hermes detection incident?

In early April 2026, Anthropic announced that third-party harnesses such as Hermes and OpenClaw would no longer be allowed to consume Claude Pro and Max subscription quotas. Anthropic's stated reason was that subscriptions were not built for the usage patterns of these tools, which can run Claude Code in loops at token volumes flat-rate plans were never priced for. Users of those tools were directed to API billing. That part was policy, disclosed and contested by developers.

The bug came later. Around April 25, a developer reported that a $200 per month Claude Code Max plan had been bypassed and a separate overage charge of more than $200 had appeared while about 86% of the plan's capacity sat unused. The trigger was the string HERMES.md in a Git commit message, not active use of Hermes. Reports said the match was case-sensitive, and that the same word in other forms did not fire. The mechanism, as described in public reporting, was that Claude Code pulls Git status into its system prompt and a harness-detection routine matched strings associated with Hermes and OpenClaw, then routed the session to extra-usage billing.

Anthropic's first support reply did not offer a refund. After the report spread, an Anthropic engineer publicly acknowledged "a bug with the third-party harness detection and how we pull git status into the system prompt" and said affected users would receive a refund plus another month of credits. The detection bug was patched, and the underlying policy on subscription quotas stayed in place. The incident timeline and the three governance dimensions it raised are in Governance Lessons of the Hermes Story.

What is a Claude harness, and how do the main tools compare?

A Claude harness is any third-party tool that wraps Anthropic's Claude model, usually through Claude Code or the Anthropic API, and adds its own interface, orchestration, or workflow layer. Seven tools come up in this conversation, and they are not interchangeable.

ToolType and builderModelsAuthenticationSurface
Claude CodeFirst-party CLI, AnthropicClaude onlyAnthropic OAuth or API keyTerminal
HermesThird-party harness, Nous ResearchMulti-providerOAuth, API key, or Claude Code credential reuseTerminal
OpenClawOpen-source Claude Code framework, communityClaude (Code)Inherits Claude Code authTerminal, scriptable
AiderAI pair-programmer CLI, Paul Gauthier and open sourceMulti-providerProvider API keyTerminal
ClineVS Code agent extension, Saoud Rizwan and open sourceMulti-providerProvider API keyVS Code
ContinueIDE extension, Continue.devMulti-providerAPI key or OAuthVS Code, JetBrains
CursorAI-first editor, Anysphere (not strictly a harness)Multi-providerCursor accountStandalone editor

Why did the detection flag Hermes and OpenClaw specifically?

In EPR's analysis, the detection targeted the two tools that most directly enable autonomous, machine-paced sessions running on a user's existing Claude Code authentication. Hermes can route through Claude Code's credential store, and OpenClaw inherits that authentication by design. Aider, Cline, Continue, and Cursor generally require their own provider credentials, which separates a user's Claude subscription quota from the tool's usage. Whether Anthropic extends detection to the usage patterns those tools enable is the open question the category is watching.

How are agent frameworks reshaping the Claude ecosystem?

Agent frameworks bet that the foundation model is becoming infrastructure and the durable product surface is the layer between the model and the user. The Hermes incident was reported as a billing story, but it is also a category story that repeats around GPT and Gemini.

For a foundation lab, the bet disintermediates the customer relationship. A developer using Claude Code is Anthropic's customer, while a developer using Hermes is Nous Research's customer, with Claude as the engine under the car and the car as the brand. Labs that do not manage the shift risk becoming wholesale utilities. Anthropic's detection layer was, in part, an attempt to preserve the direct relationship where it still exists, and the implementation was botched even though the strategic concern is real.

If labs price like infrastructure, the layer above can price like software, closer to the customer and able to switch providers. The cloud era offers the precedent: the providers sell compute at cloud-provider multiples, and companies built on top of them earn application-software multiples. On current evidence, EPR expects five things: consolidation around a handful of framework leaders, competitive first-party harnesses from the labs, further detection and policy episodes, procurement that formalizes around the category with telemetry and enforcement language in contracts, and multi-provider routing as table stakes.

What does the incident mean for communications and governance?

The incident shows that AI platforms now inspect the environments they run in and sometimes act silently on what they find. Claude Code reads Git status, other assistants remember across sessions or read workspace context, and the platform observes the surface around the prompt. The deeper risks apply across every AI platform:

  • Opaque telemetry: users do not know what AI tools read, and this one was caught only because a developer traced it through a Git history.
  • Silent enforcement: a rerouted session or surprise charge can be disputed only after the bill arrives.
  • Keyword false positives: string matching has no semantics, and a filename is not evidence of use.
  • Support escalation as the only correction path, which worked here only after the story spread.

For enterprise buyers, governance posture is replacing capability as the differentiator: what the platform reads inside their environment, what enforcement it can take, how those actions are surfaced, and what recourse exists when a false positive fires. For communications teams, that means press rooms and trust centers need a governance-posture page, crisis playbooks need a "platform did something silently" template, and buyer-facing content has to anticipate the audit by saying less and proving more.

How does Hermes connect to Amazon v. Perplexity?

Both are the same structural question on different surfaces: when a user authorizes a third-party agent, what can the underlying platform do about it? In Hermes, an AI lab detected and rerouted third-party harnesses. In Amazon v. Perplexity, Amazon sued to block Perplexity's Comet browser from acting on a logged-in user's behalf, and the case reached the Ninth Circuit in June 2026. Check the case page for current status.

Explore the Hermes cluster

Frequently Asked Questions

What is Hermes?

Hermes is a third-party Claude Code harness developed by Nous Research. It wraps Anthropic's Claude Code CLI inside a separate terminal interface with its own skills library, credential routing, and multi-turn orchestration.

Is Hermes built by Anthropic?

No. Hermes is built by Nous Research, an independent AI lab, and uses Claude through pay-per-token API access or, where authorized, Claude Code's credentials.

What was the HERMES.md billing bug?

In April 2026, Claude Code's third-party harness detection matched the string HERMES.md in Git commit messages and routed affected sessions from subscription quota to extra-usage billing, even when no harness was running.

Did Anthropic confirm the bug?

Yes. An Anthropic engineer publicly acknowledged a bug in third-party harness detection and in how Git status is pulled into the system prompt, and said affected users would get a refund plus another month of credits.

How much were users overcharged?

The most-cited case involved a $200 per month Max plan user who was charged more than $200 in overage while about 86% of the plan's capacity sat unused.

Why was Anthropic detecting third-party harnesses at all?

Anthropic's stated position is that subscriptions were not built for the usage patterns of third-party tools, which can run Claude Code in loops at token volumes flat-rate plans were not priced for.

What is the difference between Hermes and OpenClaw?

Both orchestrate Claude Code. Hermes is built by Nous Research with its own terminal and skills system, and OpenClaw is an open-source community framework that inherits Claude Code's authentication.

Will more incidents like this happen?

Likely. Detection logic, telemetry surfaces, and silent enforcement are growing faster than disclosure norms, and the conditions that produced this incident are still present.

EPR Editorial Team
Written by
EPR Editorial Team

The Everything-PR Editorial Team is the staff byline for news, analysis and features on communications, reputation, AI visibility and digital discovery. Everything-PR has published since 2009. Every article is reviewed by a human editor before publication. Coverage follows the Editorial Policy, and substantive corrections are noted on the article under the Corrections Policy.

Related reading

Other news

See all

Most brands are invisible inside AI search. Is yours?

EPR publishes the data every week.

Free. Weekly. Unsubscribe anytime.