Hermes Agent Review

Hermes Agent review, verified 17 October 2026: installation, capabilities and pricing checked against Nous Research's live pages. The judgment rests on continuity, access control and cost per accepted result. The workflows below are proposed acceptance tests, not claims of hands-on performance.

What Hermes Agent Actually Is

Hermes Agent is Nous Research's open-source software for giving a language model tools, saved context and a place to execute work. The MIT license lets you use and modify the software; you supply a model connection and a machine, or use optional hosted services. It can store conversations and memories, create reusable skill documents, and receive requests through desktop, terminal or messaging interfaces. Think of it as the assistant's working environment: the language model reasons, Hermes manages the surrounding workflow, and your configuration determines what the assistant can reach. Official Hermes Agent site.

Hermes Agent Review: Who Should Install It?

Hermes Agent fits a founder or operations lead who repeats work across sessions and wants to inspect the instructions and context behind it. A weekly customer-feedback briefing is a good candidate: collect approved exports, group complaints, link evidence and prepare a draft for the product meeting. An autonomous refund agent with unrestricted billing access is a poor starting point.

The shortlist turns on the job you want to delegate:

The alternative roles come from the official Claude Code overview, Codex CLI documentation and OpenClaw documentation. These are fit judgments, not a ranking of task-completion accuracy.

Hermes Agent vs Claude Code

Claude Code is my recommendation when the desired outcome is a reviewed code change. Hermes earns its installation when the work continues beyond a repository: keeping an operator's preferences, reusing a reporting procedure and delivering the result through a configured chat channel. Choose Codex on the same principle if it already fits your development workflow. There is little value in maintaining another agent simply to ask the same coding questions.

For an existing OpenClaw installation, require a specific workflow improvement before migrating. The separate Hermes Agent vs OpenClaw comparison covers that decision in detail.

Who should skip Hermes: a nontechnical owner without someone responsible for updates and credentials; a developer whose current coding assistant already handles the work; and any operator who cannot check whether an automated result is correct. In the last case, keep a human approval step in your existing business application before adopting either self-hosted assistant.

Installation: Desktop First, Terminal When You Need It

Hermes Desktop is the easiest entry point for a founder who wants a graphical interface. The official site provides installers for macOS and Windows. Desktop uses the same configuration, keys, sessions, memory and skills as the command-line agent. You are choosing an interface, not starting a separate assistant. Desktop documentation.

For Linux, macOS or Windows through WSL2, the documented terminal installer is:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

Native Windows also has a PowerShell path:

iex (irm https://hermes-agent.nousresearch.com/install.ps1)

These commands download and execute installation code. Use the official domain and review the script first if your business requires that. The source installer prepares dependencies and browser/computer-use tooling, so a successful download is only the beginning of setup. Follow the installation guide, then run hermes model to choose your provider. For Nous Portal's combined model and hosted-tool setup, the documented command is hermes setup --portal.

Start with a disposable working folder and a copy of an export. Your first task should prove Hermes can produce a useful file before you attach business accounts.

Persistent Memory: Keep Business Context Between Sessions

Hermes Agent's memory is useful for stable working preferences: what a report means, where approved inputs live, and how the owner wants the answer presented. It is a poor substitute for the underlying business records.

The built-in system stores environment and workflow notes in MEMORY.md, and user preferences in USER.md. It loads a snapshot when a session begins and can retrieve older conversations through session search. A new memory write reaches disk, but the session's initial snapshot stays fixed. Persistent memory documentation.

For the feedback briefing, save a preference such as “separate requests from confirmed bugs.” Keep the current customer counts in the source export, where they can be recalculated. Otherwise yesterday's number can become today's remembered assumption.

The acceptance test is straightforward: ask Hermes to save the preference, inspect the saved entry, then start a new session and give it a fresh export. Check that the classification follows the preference without carrying over last week's counts. A message saying it remembered something is not proof of a successful write.

Use /new when a task ends. A messaging conversation can survive a gateway restart, so restarting the process is not the same as starting fresh. If you want to approve what becomes durable context, enable memory.write_approval; messaging and background saves then wait for review through /memory pending.

The payoff: less repeated briefing. The wall: retained context can preserve a mistaken assumption just as effectively as a useful one.

Self-Created Skills: Turn a Good Run Into a Repeatable Procedure

Hermes Agent's strongest reason to stay installed is its ability to turn a useful method into a skill you can inspect and reuse. A skill is a procedural document, with supporting resources when needed, that the agent loads on demand. Creating one changes the instructions available to future runs; it does not establish that the underlying model has been retrained. Skills documentation.

A useful feedback-report skill would specify the input columns, the classification rules, how to preserve evidence links, what counts as a duplicate, and when to stop for missing data. “Write a great report” is not a procedure. “Reject a file without the customer identifier and report the missing field” is.

The approval gate also applies to changes from the background learning process. That matters once the reporting skill works: a later conversation should not quietly redefine the company's categories. Review edits as changes to an operating procedure, including additions that appear helpful.

A founder should judge the learning loop by whether the next accepted report needs fewer corrections. A growing folder of skill files is activity, not evidence of improvement.

Messaging and Schedules: Put the Briefing Where Work Happens

Hermes Agent connects to Telegram, Discord and Slack through its messaging gateway, the running service that accepts messages and delivers responses. The attraction is practical: an operator can request a draft from the place they already coordinate work. Each platform still needs its own credentials and configuration. Messaging gateway documentation.

For the weekly briefing, start with your own account and a private destination. Configure the gateway with hermes gateway setup, then verify a request and its response arrive in the intended place. Deliver drafts there before considering a wider audience. Access to the bot is access to whatever work its tools permit; a channel invitation should not silently become operating authority.

Hermes supports recurring jobs, attached skills and delivery to configured destinations. A sensible schedule asks it to summarize the latest approved export and return a draft with evidence links. The machine and gateway must remain available for the work to run. A sleeping or stopped local installation is not an always-on service. Scheduling documentation.

Pin the scheduled job's model once the workflow is accepted. The documented selection order is the job's pinned model, then the cron-specific model, then the main model selected when the job fires. Without a pin or cron default, changing the model for an unrelated chat can also change the next report's model. That affects both repeatability and spending.

If the procedure eventually becomes deterministic, Hermes also documents script-only scheduled jobs without a model call. Use reasoning where it adds value; a fixed export transformation may be better as a script.

Models and Execution: Choose the Brain and Its Workspace

Hermes Agent lets you separate the model you pay for from the environment where commands execute. That flexibility is valuable when a business wants to change providers without rebuilding every operating instruction.

The documented options include Anthropic's Claude, OpenAI models, Google Gemini, providers through OpenRouter or Nous Portal, and compatible local endpoints such as Ollama or vLLM. Use hermes model to choose the route. Model availability is a provider decision, and a successful connection does not prove the model handles your tools reliably. Provider documentation.

For the briefing, compare a candidate model against the same approved input and acceptance standard. Check evidence preservation, arithmetic, tool execution and the missing-column test. Count retries and correction time. A cheaper response that produces an unusable report is not the cheaper outcome.

Execution is a separate choice. Hermes can run terminal commands locally, in a container or through a remote backend. A container is an isolated execution environment; it does not make a remote language-model request local. If your report includes confidential records, decide which fields the model provider may receive as well as which folders the agent may open. Execution configuration.

Hermes Agent Pricing: What an Operator Should Budget

Hermes Agent's software is free. For a small pilot, bring an existing API key and pay that provider for usage, or choose Nous Portal if you want models and hosted tools on one balance. The Portal's Free tier is limited to free models and includes no monthly credits; Plus is the starting paid option I would consider for this reader.

These are the subscription figures shown on Nous Portal's own page, checked 17 October 2026:

Portal lists standard rate limits for Free, and high rate limits plus hosted tool usage for paid tiers. Hosted tools bill against the same credits as models. Pick Plus when that combined setup is useful; move higher only after observing sustained consumption. Buying the larger tier does not establish a more reliable reporting process.

For a concrete budget illustration, Portal lists Google Gemini 2.5 Flash at $0.30 per million input tokens and $2.50 per million output tokens. Tokens are the units a model uses to process text. Assume an accepted report consumes 100,000 input tokens and 10,000 output tokens across all its calls: the inference charge is (0.1 × $0.30) + (0.01 × $2.50) = $0.055. Twenty such reports would consume $1.10 of inference at those listed rates. These are hypothetical usage inputs, not measured Hermes results; the calculation excludes tool calls, hosting, retries and cache discounts. Portal model catalog.

That example is a reason to measure before subscribing heavily. Your useful metric is total spend on all attempts divided by reports you accept, with human correction time recorded alongside it.

Hermes Cloud adds a separate hosting bill. Nous documents running charges for compute plus storage, and storage-only charges while stopped, billed from Portal credit. Inference and tools are additional. The current instance sizes and prices are shown in the deployment flow, so do not read a Portal subscription as an included always-on server. Hermes Cloud billing.

Hermes Agent With Claude Code Subscription

Hermes Agent can use Claude through an Anthropic API key, independently billed from a Claude subscription. Its current provider docs also describe an OAuth route requiring Claude Max plus purchased extra-usage credits; they say Hermes uses those extra credits, not the base Max allowance, and that Claude Pro cannot use that route. Treat this as the documented Hermes integration condition, not a promise that any existing Claude plan covers the work. The API-key route is the clearer budget boundary. Hermes Anthropic setup.

Limitations: Where Hermes Agent Breaks the Workflow

Hermes Agent's practical limits are the boundaries between saved knowledge, successful execution and permitted action. The following failures matter more to a business than whether the agent produces an impressive demonstration.

A remembered instruction can be wrong or never saved. The memory mechanism needs a successful write, and a new session needs the correct profile. Keep records in their source system, verify important saved preferences and reject results that rely on unsupported recollection. Memory improves continuity; it does not make the agent a reliable system of record.

A saved skill can repeat a bad method. A model can turn an accidental workaround into reusable instructions. Require a missing-input test and inspect changes to procedures that affect customer-facing work. Self-created skills are a maintenance surface, not a substitute for quality control.

Scheduling still needs an operator. A job can be interrupted, lose an account connection or produce output that fails review. Hermes' cron docs describe interrupted executions that remain paused for operator review. Decide who notices a missing briefing and who can resume it. A scheduled job existing in a dashboard is not evidence that its result arrived. Cron recovery behavior.

The sandbox label can hide a broader access grant. The local backend runs on the host without container isolation. Docker and cloud backends provide an execution boundary, but a credential deliberately passed inside can still be read and used there. Hermes skips dangerous-command checks for several isolated backends because the container is the boundary. Inspect mounted files and account permissions before assuming every consequential action will ask you first. Security boundaries.

Scope the Machine, Accounts and People Separately

For a business pilot, configure the job around the minimum access needed:

1. Use a separate working environment with copied inputs. On the local backend, assume commands can reach resources available to the operating-system user, subject to the configured controls.
2. If using Docker, install and verify Docker first, select that backend, and expose only the necessary folders. Do not mount an entire home directory merely to make file access convenient.
3. Give connected services a dedicated account or credential with limited permissions. Start with reading and drafting. Keep sending, deleting and billing changes outside the pilot's authority.
4. Restrict who can command the gateway through explicit allowlists or approved pairing. Leave global and platform-specific allow-all options disabled.
5. Review retained memories and skills, and keep a backup of the Hermes data directory. Local storage does not prevent selected context from reaching a remote model.

The following documented settings are useful starting controls, not a complete security configuration:

hermes config set terminal.backend docker
hermes config set approvals.mode manual
hermes config set approvals.cron_mode deny
hermes config set memory.write_approval true
hermes config set skills.write_approval true

The manual setting concerns flagged dangerous commands where the approval layer applies; it does not require approval for every action, and it does not reinstate checks skipped by an isolated backend. Memory and skill approvals control what becomes persistent knowledge. Configuration commands, approval behavior.

Verdict: Install for Repeated Work You Can Verify

Hermes Agent is a strong candidate when the same business procedure returns often enough that preserved context and inspected skills reduce repeated setup. It is a poor purchase of your time when the job is mostly coding, the workflow changes constantly, or nobody can maintain its access and judge its output.

The decision rule: keep Hermes only if a representative recurring task produces an acceptable result with less total correction and maintenance effort than your current method, at an acceptable combined model, tool and hosting cost. Otherwise keep the existing process or coding assistant.

On Monday, choose one internal briefing. Supply a copied export, define the output standard and failure conditions, and run it under restricted access. Inspect the result, memory and proposed skill. Only after those pass should you schedule it or connect another account. That sequence makes the business case observable.

How to Uninstall Hermes Agent

Hermes Desktop documents hermes uninstall for a self-managed installation that removes the app and agent while retaining data. hermes uninstall --gui removes only the graphical interface; hermes uninstall --full also removes user data. Back up what you need before choosing the full removal. Package-managed or externally owned installations should be removed through their package manager or operating system. Official uninstall instructions.

Hermes Agent FAQ

Use the AI business workflow audit checklist to choose the first workflow worth delegating.