If I already pay for ChatGPT or Claude, I do not need Hermes Agent just to get another AI that can think, write code or edit files. Claude Code and Codex are already extremely capable agents. The reason to run Hermes is different: I want an agent runtime that remains mine when I close the chat window.
That distinction matters more now than it did six months ago.
Hermes v0.20 makes the overlap impossible to ignore
Hermes Agent v0.20.0 shipped on August 3, 2026 as the Herald Release.1 It is a much better release to use for this comparison than the earlier v0.2 because the project has moved well beyond being an interesting terminal agent.
Some of the additions are immediately visible: real-time conversational voice, voice-note handling across messaging platforms, grounded research citations, Agent-to-Agent protocol support, signed outbound webhooks, a much richer desktop application and a wave of CLI improvements.1
But the feature I find most revealing is much less flashy:
hermes import-agent
Hermes can now inspect an existing Claude Code or OpenAI Codex CLI setup and migrate much of it into Hermes. For Claude Code, that includes CLAUDE.md, command permissions, MCP servers and skills. For Codex, it can import AGENTS.md, MCP configuration, memories and skills. Credentials are deliberately excluded.2
That tells me something about where Nous Research thinks Hermes belongs.
It is no longer positioning itself merely as another environment in which I teach an AI how to work. It is increasingly positioning itself as a runtime capable of hosting the conventions I already built elsewhere.
The Claude Code skill I spent time writing does not necessarily have to be abandoned.
The MCP servers do not necessarily have to be recreated.
The project instructions do not necessarily have to be rewritten from scratch.
Hermes is effectively saying: bring them with you.
That is a much more interesting proposition than “our coding agent has more tools.”
Claude Code and Codex are excellent workers, not complete operations
Claude Code and Codex are strongest when the centre of the problem is a codebase. They can read a repository, apply changes, run tests, use project instructions, make use of skills, and work with subagents or task automation. Anthropic documents hooks and skills for Claude Code; OpenAI documents skills, subagents, sandboxed agents and scheduled Codex automations. 4 5
That is substantial capability. A software team with a well-maintained repository should usually start there for programming work. Claude Code or Codex has a sharper default for editing, testing and reviewing code than a general-purpose business agent. They are also easier to justify if the only repeated work is inside a Git repository.
But a business process is rarely only a repository. A lead needs a follow-up, an operations report needs data from a spreadsheet and a SaaS dashboard, an account manager needs a reminder in the right channel, and a decision needs to be recorded so that next month’s report follows the same rules. Claude Code and Codex can participate in those workflows, but they do not by themselves decide the organisation’s memory policy, evidence trail, approval boundary or messaging delivery model.
That is where Hermes becomes useful. Its value is not that it can write code more impressively than Claude Code or Codex. Its value is that it can give the code worker a durable place in a wider system.
The comparison is about the unit of work
| Question | ChatGPT / Claude | Claude Code / Codex | Hermes Agent |
|---|---|---|---|
| Best unit of work | A conversation, document, analysis or one-off task | A repository change or engineering task | A repeatable cross-tool operating process |
| Default context | The active chat and uploaded material | The active repository and its instructions | Sessions, durable memory, skills, project context and connected tools 3 |
| Skills and reuse | Product-specific prompts and capabilities | Coding-oriented skills and project instructions 4 5 | Procedural skills intended to be created, reused and improved across sessions 3 |
| Hooks and events | Limited to each product’s own integrations | Claude Code hooks; Codex automation and API/webhook surfaces 4 5 | Hooks, MCP, gateway channels, schedules and tool execution in one agent environment 3 |
| Artifacts | Strong for presenting chat output | Strong for files and code deliverables | Desktop artifacts with live preview; can also produce operational files and messages 2 |
| Best reason not to use it | You need work to run beyond the chat | The task is not principally software delivery | You do not have a repeatable process worth maintaining |
The table is not a claim that Hermes replaces the others. It is an argument for composition. Use ChatGPT or Claude for thinking and drafting. Use Claude Code or Codex for a bounded engineering task. Use Hermes when the result must flow through a recurring process with durable context, controls and delivery.
SMEs need less novelty and more continuity
For an SME, the practical test is simple: how many useful decisions are made repeatedly, then lost in chat histories, inboxes and staff memory?
Consider a small services firm. Each week it may need to review new leads, identify stalled opportunities, prepare client follow-ups, collect market changes, update a project record and send a director a short exception report. None of those tasks requires an exotic model. The difficulty is continuity: using the same definitions, retrieving the prior record, not sending the same reminder twice, keeping confidential information out of the wrong channel, and knowing who can approve an external action.
A Hermes-style system can be appropriate when those rules are explicit. A skills file can describe how the weekly review is assembled. Memory can preserve the customer’s preferred reporting format. A scheduler can run the draft at a known time. A human can retain approval for outbound communication or financial decisions. The work becomes inspectable rather than magical. This is the kind of operations automation that an SME can actually govern.
The counter-case matters. If the company does not have a stable process, Hermes will preserve confusion faster. If the inputs are messy, data access is unclear or no one owns the exceptions, an agent adds a new control surface without solving the underlying problem. A ChatGPT or Claude conversation may be the better first move while the team clarifies the process manually.
Do not buy a second brain when the business first needs a shared checklist.
Start with one process that deserves to exist
I would not begin by connecting every inbox, spreadsheet and chat channel. Pick one workflow with a clear cadence and measurable failure mode: lead follow-up, weekly pipeline review, renewal tracking, project status reporting or document intake.
Write down the input, the person who owns it, the actions that require approval, the record that counts as truth and the definition of done. Then decide whether a model chat, a coding agent or an agent system is the narrowest tool that can carry it.
If a process ends when the chat ends, ChatGPT or Claude may be all that is needed. If it ends in a pull request, Claude Code or Codex is likely the better first tool. If it has to happen again next week with the same rules, records and channels, Hermes is no longer redundant. It is the layer that turns intelligence into operations.