Hermes Agent and Nous Portal: What the MIT License Does Not Cover

Last updated: August 19, 2026

Hermes Agent is MIT-licensed software from Nous Research that runs as a long-lived agent process on infrastructure the operator controls. Nous Portal is a separate commercial service from the same company, with free and paid tiers, that supplies models, hosted tools, billing, and optional cloud hosting to that agent. The MIT license governs the repository. Portal adds its own service terms, privacy policy, and data grants.

Verdict: Portal is the fastest documented setup path, and any organization handling confidential, regulated, or client-controlled material should resolve Section 12.3 and the Privacy Mode timing limit before routing the first payload through it.

What the MIT license covers

The Hermes Agent repository ships under the MIT license. That grant covers use, copying, modification, distribution, sublicensing, and sale of the software, on condition that the copyright and permission notice are retained, and it supplies the software without warranty. No Nous Research software fee attaches to it.

The MIT grant covers code. It says nothing about prompts, outputs, credits, or traffic sent to any hosted service, because it is not a service agreement and was never meant to be one.

Portal is optional in the product sense. The repository documentation states that the agent works with Nous Portal, OpenRouter, OpenAI, or an operator-supplied endpoint, and that providers are switched with a single command. Portal is presented as the convenience route, not the required one.

That distinction should not be compressed into “free versus contract.” MIT is itself a legal license, and a direct-provider deployment creates its own relationships with whichever model, hosting, and tool vendors the operator selects. The narrower and defensible statement is that enabling Portal adds a second set of terms on top of the repository license.

Sources: NousResearch/hermes-agent LICENSE, accessed 19 August 2026 · NousResearch/hermes-agent README, accessed 19 August 2026

Where the two documents disagree

Nous Research publishes two governing documents for its services, and they state different scopes.

The Terms of Service define the governed “Platform” by enumeration: the Nous Portal application, the Nous Inference API, the Nous Forge application, the Nous Research chat interface at hermes.nousresearch.com, and the Nous Research merchandise store. The self-hosted GitHub repository is not named in that list.

Section 11 of those same Terms incorporates the Privacy Policy by reference. The Privacy Policy sets out its own scope, and that scope is written more widely. It says it describes information Nous Research may collect or process during use of its websites and tools, listed as including the Nous Portal platform, “Hermes and other AI models, agents, and inference tools made available by Nous,” Nous APIs and developer tools, and Nous websites and related subdomains.

Whether an MIT-licensed agent that Nous Research distributes on GitHub falls inside the phrase “agents … made available by Nous” is not resolved by either document. FSR is not offering a legal interpretation of it, and this briefing does not claim the two scopes are inconsistent as a matter of law.

The buyer consequence is operational. A local-model or direct-provider configuration avoids the documented Portal account, the Nous inference route, the Tool Gateway, and the credit ledger. That much follows from the product documentation. What does not follow from the published documents is a clear answer to whether the privacy scope reaches a configuration in which no Nous service is contacted. An organization that needs that answer should obtain it in writing rather than infer it from the license file.

Sources: Nous Research, Terms of Service, 18 June 2026 · Nous Research, Privacy Policy, 11 June 2026

What changes when Portal becomes the provider

The repository documentation describes a single onboarding command, hermes setup --portal, and states what it configures: an OAuth login to Portal, Nous set as the inference provider, and the Tool Gateway enabled for hosted web search, image generation, text to speech, and a cloud browser. That is published setup behavior, not an FSR observation.

The Terms describe acceptance separately. They state that the Terms are accepted as soon as the person has received confirmation of account creation and the credentials needed to log in, or, for services and parts of the site that do not require an account, upon gaining access to those services.

Read together, the published documents show that a configuration step and a contract acceptance step occupy the same part of the workflow. They do not show what the operator is presented with while it happens.

FSR cannot currently establish, and this briefing does not assert, any of the following: whether the Terms and Privacy Policy are linked beside the signup control, whether a separate acceptance checkbox appears, whether the OAuth authorization screen summarizes data use, whether Privacy Mode is offered before the first inference request, or whether free and paid signup use the same sequence. Those require a dated capture from a fresh account, which is outside Tier C scope.

Sources: NousResearch/hermes-agent README, accessed 19 August 2026 · Nous Research, Terms of Service, 18 June 2026

Section 12.3: the two-stage data grant

Section 12.3 is the clause most likely to decide whether an organization can use Portal. It contains two distinct grants, and the second one is the one that matters for confidential material.

The first grant is a service and development license. The customer grants Nous Research a non-exclusive, worldwide, royalty-free, fully paid-up, transferable, and sublicensable license to use, reproduce, modify, create derivative works from, and otherwise process customer data, for purposes that include operating the services, developing new products, and training, fine-tuning, evaluating, and enhancing machine learning models.

The second grant sits in subsection (c). That subsection first defines “Service Data” to include data, content, inputs, prompts, outputs, results, usage data, and metadata processed through the services. It then addresses aggregated, anonymized, de-identified, or derivative material, which Nous Research may make available to third parties without restriction. Immediately after that, it states a further right in its own sentence: the customer grants Nous Research the right to provide, license, or otherwise make available Your Data and Service Data to third parties for any lawful purpose, including for those third parties’ own use, analysis, and product development.

That final sentence is not limited by the aggregation and de-identification language that precedes it, because the aggregation case is handled in a separate sentence of the same subsection.

This briefing does not benchmark that wording against competing service agreements, and makes no claim about whether it is unusual. It also does not allege that Nous Research currently distributes any particular customer prompt, and the contract does not indicate how often the right is exercised. The point is narrower and sufficient: a permission that a data owner cannot grant should not be granted by an engineer during a setup command. Organizations bound by client confidentiality, NDA terms, or regulated data obligations should route this clause to whoever signs off on data handling before enabling Portal.

Sources: Nous Research, Terms of Service, Section 12.3, 18 June 2026

What Privacy Mode does and does not reach

The Privacy Policy documents a control called Privacy Mode, reached by logging in, opening Account Settings, and selecting it under Privacy Settings.

When enabled, the policy says Nous Research will not store inference payloads and will not use them for training, product improvement, or support. Inference payloads are described as prompts, inputs, uploads, messages, outputs, and responses. Three limits are stated in the same passage.

Exceptions remain, to the limited extent needed to maintain security, integrity, and operation of the services, comply with law, or detect and investigate fraud, abuse, or security incidents. Operational metadata is outside the control, and the policy states that Privacy Mode does not prevent collection, storage, or use of that metadata for routing, securing, operating, debugging, or billing. And the timing is one-directional: the policy states Privacy Mode may apply only on a prospective basis and may not affect information collected, processed, incorporated into models, or otherwise used beforehand.

The practical consequence is sequencing. Enabling Privacy Mode after an evaluation period does not, on the policy’s own terms, reverse what happened during that period. Any organization intending to use it should enable it before the first payload rather than after a pilot.

The policy separately notes that some services use third-party model and infrastructure providers, and that prompts, inputs, and outputs may be shared with those providers, who process the material under their own terms.

Sources: Nous Research, Privacy Policy, 11 June 2026

What the money buys, and what is undefined

Portal publishes four tiers. Free carries $0 in monthly credits with access to free models only and standard rate limits. Plus is $20 per month for $22 in monthly credits with a $10 rollover cap. Super is $100 for $110 with a $50 cap. Ultra is $200 for $220 with a $100 cap. The three paid tiers add access to hundreds of models, hosted tool usage, and higher rate limits. A top-up panel accepts custom amounts through Stripe.

The paid plans therefore combine recurring credits with plan-level rate and feature benefits. Credits are consumed by model inference and hosted tools, which are separately priced per unit for services including Firecrawl search credits, Browser Use session minutes and proxy bandwidth, FAL and Krea image generation, Modal sandbox compute, and OpenAI audio.

Hermes Cloud, the hosted agent instance, draws on the same balance and bills daily. The published rates are $0.29 per day running and $0.03 stopped for a small instance, $0.56 and $0.03 for medium, and $1.09 and $0.03 for large. The information page states that these prices exclude inference and tool usage, and that a minimum $2 credit balance is needed to start.

The pricing page names a rollover cap as a plan attribute for each paid tier. FSR did not locate a definition of that term, or any statement of credit expiry, in the Terms of Service. Section 5.4 covers fees and refunds without addressing credit balances. What happens to unused credits above the cap, on downgrade, or on cancellation is therefore not established by the documents read here, and buyers should confirm it against the checkout, help, or account billing surfaces before relying on any assumption.

Sources: Nous Portal pricing, accessed 19 August 2026 · Nous Portal product information, accessed 19 August 2026 · Nous Research, Terms of Service, 18 June 2026

Cancellation, deletion, and surviving rights

Three provisions govern what happens at the end of the relationship, and they need to be read together rather than separately.

On termination, the Terms state that Nous Research will store customer data for six months so the account can be reopened or the data exported, then permanently delete the account within six months of the effective date. A customer who specifically requests earlier deletion is to have that request fulfilled within one month.

On downgrade rather than termination, Section 5.5 warns that moving to a cheaper plan may cause loss of features, capacity, and customer data. It does not specify which data classes are affected or whether the six-month export window applies in the same way.

On survival, Section 17.3 lists the provisions that continue after termination, and Section 12 is among them. Account deletion and the continuation of intellectual property and data provisions are separate questions, and the Terms treat them separately.

On the agent side, the documentation places the agent’s skills in ~/.hermes/skills/ on the operator’s own machine and states that skills follow the agentskills.io format, which may reduce migration friction to other compatible runtimes. Backup, profile export, and session export commands are documented, and their exact contents and behavior were not verified by FSR for this briefing. Anyone relying on a full backup should note that it is documented as including credentials and should be secured accordingly.

Sources: Nous Research, Terms of Service, Sections 5.5, 9.2, 17.3, 18 June 2026 · Hermes Agent Skills System documentation, accessed 19 August 2026

Questions to resolve before procurement

Two deployment settings are worth naming before the contract questions, because they are documented, they are defaults, and they are cheap to change.

The skills documentation states that by default the agent writes skill files freely, including from the background self-improvement review that runs after a turn, and that a gate exists at skills.write_approval with false as the documented default. Memory writes carry an equivalent gate. Separately, the project’s security policy states that the only security boundary against an adversarial language model is the operating system, that in-process approvals and scanners are heuristics rather than containment, and that operators running the default local backend against untrusted input surfaces are outside the supported security posture. The quickstart recommends Docker or a remote server instead.

For a procurement reviewer, the shortest useful list of written questions to send Nous Research is: whether Section 12.3’s third-party availability grant can be amended or waived for a paying customer; whether Privacy Mode changes the operation of that grant; how credits behave on rollover, downgrade, and cancellation; whether a standalone data processing addendum, subprocessor list, and residency commitment exist; and whether the Privacy Policy scope reaches purely local, non-Portal deployments.

Sources: Hermes Agent Skills System documentation, accessed 19 August 2026 · Hermes Agent Security Policy, accessed 19 August 2026 · Nous Research, Terms of Service, 18 June 2026

FAQ

Methodology and evidence limits

This is a Tier C briefing. Future Stack Reviews did not create a Nous Portal account, purchase credits, install or run Hermes Agent, enable Privacy Mode, view an account dashboard, or test cancellation and export. No figure or behavior described above is a measurement.

Documents read for this briefing, all on 19 August 2026 unless otherwise dated: the Nous Portal Terms of Service dated 18 June 2026; the Nous Research Privacy Policy dated 11 June 2026; the public Nous Portal pricing and product information pages; the NousResearch/hermes-agent LICENSE, README, SECURITY.md, and Skills System documentation.

Four limits apply, and each one narrows a claim above.

First, the scope relationship between the Terms and the incorporated Privacy Policy is unresolved in the published text, and this briefing records it as unresolved rather than resolving it.

Second, everything about the signup and OAuth presentation is untested. The published documentation shows what the setup command configures. It does not show what a person sees in a browser during account creation.

Third, credit rollover and expiry mechanics are named on the pricing page but not defined in the documents read here.

Fourth, several statements above are absence findings. Where this briefing says an item was not located, that means FSR did not find it in the public sources checked, which is not the same as establishing that it does not exist.

Prices, plan structures, contract text, and release versions on this platform change without notice. The Terms themselves provide that changes take effect on continued use or 30 days after posting.

Verdict

Choose the data path before configuring the provider, not after.

For public repositories, open data, personal projects, and low-sensitivity experimentation, Portal is a reasonable default. One login replaces several API keys, the free tier allows evaluation at no cost, and the published clauses are visible to anyone who reads them.

For client-confidential material, regulated personal data, NDA-controlled documents, or any workload where a third party owns the data, the sequence runs the other way. Establish whether the organization can grant the rights in Section 12.3. Obtain any required processing addendum, subprocessor list, and residency commitment in writing. Enable Privacy Mode before the first production payload rather than after a pilot. Until those are settled, keep the workload on a local model or a separately contracted provider, which the agent supports by design.

The one inference this briefing asks readers not to make is the common one: that an MIT badge on a repository says anything about what a hosted provider may do with data sent through its service. It does not. Those are two documents, and only one of them is a service agreement.