Terms of Use

Agreement to these terms

These Terms of Use ("Terms") are a binding agreement between you and Cymela. They apply when you visit this website, install or run the Cymela CLI, or download and use any model, dataset or other artifact we publish (together, the "Services").

By doing any of those things you accept these Terms. If you do not accept them, do not install or use the Services. Please also read our Privacy Policy, which forms part of this agreement.

Who we are

Cymela is an independent artificial-intelligence research project operated by a private individual resident in the Republic of Serbia. It is not a registered company, and nothing on this website should be read as implying that it is. Where these Terms say "we", "us" or "Cymela", they mean that operator.

Operator contact: contact@cymela.com. This address reaches the operator directly and is monitored.

The operator is not named individually on this page. Cymela is a non-commercial project that takes no payments and holds no user accounts. The operator's full legal name and address will be provided on any legitimate request — including from a court or a person pursuing a genuine legal claim — and will be published here if Cymela is incorporated or begins charging for the Services.

What we provide

The Services currently consist of:

• This website — informational material about the project's research.
• The Cymela CLI — a terminal coding agent, distributed free of charge, which runs on your own computer and connects to a third-party model provider of your choosing.
• Published research artifacts — any model weights, datasets or documentation we release, typically through third-party platforms such as Hugging Face or GitHub.

All of it is experimental research software under active development. There is no service-level commitment, no guaranteed uptime, no guaranteed support, and no promise that any feature will continue to exist, keep working, or remain free. We may change, suspend or discontinue any part of the Services at any time.

The Services are provided free of charge. You are, however, responsible for any costs you incur with third parties — most obviously the API usage fees charged by whichever model provider you choose.

Eligibility

You must be at least 18 years old, or the age of legal majority where you live, to use the Services. If you are younger, you may use them only with the involvement and consent of a parent or legal guardian who accepts these Terms on your behalf.

You must also not be barred from using the Services under the laws of your country, and you must not be located in a jurisdiction subject to sanctions or export restrictions that would make your use of them unlawful.

License

Subject to your compliance with these Terms, we grant you a personal, worldwide, royalty-free, non-exclusive, revocable license to install and use the Cymela CLI for any lawful purpose, including commercial work.

Where a specific software license file (for example a LICENSE file distributed with the package or repository) accompanies a component of the Services, that license governs your rights in that component's source code, and prevails over this section to the extent of any conflict. Published model weights may carry their own separate license, stated in the repository where they are published; read it before use, as it may impose additional restrictions.

You may not remove or obscure any proprietary notices, nor represent modified versions of the Services as official Cymela releases.

Acceptable use

This is the part that matters most. You agree that you will not use the Services — nor permit anyone else to use them through you — to do any of the following:

Unlawful and harmful activity

• Anything that breaks the law in your jurisdiction, ours, or the jurisdiction of any third party you interact with.
• Developing, distributing or deploying malware, ransomware, worms, exploits, botnets or other malicious code.
• Gaining or attempting to gain unauthorised access to any computer system, network, account or data — including penetration testing against systems you do not own or have documented written permission to test.
• Denial-of-service attacks, credential stuffing, mass scraping in violation of a site's terms, or evading rate limits, access controls or bot detection.
• Fraud, phishing, impersonation, social engineering, or the creation of deceptive content intended to mislead people about who or what they are dealing with.
• Generating, storing or transmitting child sexual abuse material, non-consensual intimate imagery, or content that sexualises minors in any form.
• Harassment, stalking, threats, incitement to violence, or content promoting self-harm.
• Producing material designed to assist in the creation of weapons — including chemical, biological, radiological, nuclear or conventional weapons — or in attacks on critical infrastructure.
• Unlawful surveillance, or processing another person's personal data without a lawful basis.
• Infringing anyone's copyright, trademark, trade secret, patent or other intellectual property rights.

Circumvention

• Disabling, bypassing, tampering with or otherwise defeating any safety feature, permission prompt, deny rule or plan-mode restriction in the CLI in order to accomplish something these Terms prohibit.
• Using the Services to circumvent the safety measures, content policies, usage limits or terms of service of any model provider or other third-party service.
• Using prompt injection, jailbreaking or any similar technique against a model in order to elicit output that these Terms, or that model provider's own policies, forbid.
• Impersonating the CLI itself — sending, or planting in any file, page or output the agent may read, text wearing the reserved AUTOMATED MESSAGE marker described in Section 09, or otherwise attempting to have your instructions reach the model dressed as ours.

High-risk uses

You must not rely on the Services, without qualified human review, in any situation where an error could lead to death, personal injury, or severe environmental, financial or infrastructural damage. This expressly includes medical diagnosis or treatment, legal or financial advice given to others, safety-critical engineering, aviation, transport control, weapons systems, emergency services and critical infrastructure. The Services are a developer productivity tool and are not certified, validated or fit for any of those purposes.

You choose the model, and you accept what it does

The Cymela CLI does not contain a language model. It is a client for one. To be exact about what you are looking at while the tool is running, what appears on your screen comes from three different places:

• The model you selected — the reasoning, the written replies, and the decisions about which tool to call and with what arguments. This is generated by a third party on their infrastructure, reached through an account and API key belonging to you.
• The CLI itself — the interface, status lines, menus, syntax highlighting, diffs, autocompletion, slash-command hints, progress indicators, formatting, error messages and diagnostics. This is our software running locally, and no model is involved in producing it.
• Your own system — the contents of files the agent read and the output of commands it ran, relayed to the screen as they were received. This originates on your machine, or from whatever your commands contacted; neither we nor the model authored it.

All three are covered by the disclaimers in Sections 12 and 13, and none of them is warranted. Nothing in this section narrows those disclaimers to model output alone. We do not warrant that the interface renders correctly, that a diff, diagnostic or token count is accurate, that relayed output is faithful or safe to act upon, or that any part of the software is free of defects. Where the CLI performs work itself rather than delegating it to the model — and it performs a great deal of it — that work carries exactly the same disclaimers.

This has a further consequence you need to understand clearly. The safety behavior of the system you are running is overwhelmingly a property of the model you selected, not of our software. Models differ enormously. Some are heavily safety-trained and refuse harmful requests reliably. Others — particularly smaller, older, open-weight or deliberately "uncensored" models — refuse far less, hallucinate more, follow injected instructions more readily, and can be induced to produce dangerous, illegal or simply wrong output with very little effort.

When you choose which model to point the CLI at, you are choosing that risk profile, and you accept it. We do not train, host, tune, filter or moderate these models, and we have no technical ability to prevent one from producing a given output. We make no representation that any model available through the CLI will refuse any particular request.

Model output can also be confidently wrong. Do not treat it as accurate, complete, secure, license-clean or fit for purpose without checking it yourself. This applies with particular force to generated code that touches authentication, cryptography, payments, personal data or anything exposed to a network.

The CLI reads, edits and executes code on your computer

This is the single most important operational risk of using the tool, and it deserves to be stated plainly rather than buried.

You accept that, and you agree that it is your responsibility to:

1. Read what the agent proposes before you approve it. The permission prompt exists so that a human decides — approving without reading defeats the entire design.
2. Keep your work in version control, and keep backups you can actually restore from.
3. Run the CLI with the least privilege that gets the job done, and not as an administrator or root user unless genuinely necessary.
4. Avoid pointing it at directories containing credentials, private keys or sensitive personal data that you do not want sent to your model provider.
5. Treat any repository, file or web page the agent reads as untrusted input, since content inside it may attempt to influence the agent's behavior.

What our guardrails actually do

The CLI includes real protective features: a permission engine that gates tool calls, configurable deny rules, a plan mode that blocks workspace mutations, a shell-command classifier, and a trust model that prevents a cloned repository from silently registering commands to run on your machine. We build these deliberately and we take them seriously.

One boundary in that last one is worth stating exactly, because the obvious reading is wider than the truth. The trust model gates execution, not reading. A cloned repository cannot run anything — no hook, no registered command, no workspace mutation — without your approval. But project instruction files such as CYMELA.md and SKILLS.md are loaded into the model's prompt automatically, with no trust gate. Text in a repository you clone can therefore influence what the agent says and proposes before you have approved anything. Treat a cloned repository's instruction files the way you would treat its code.

We also want to be honest about their limits, because overstating them would be worse than useless to you:

Automated messages, and the marker that labels them

The CLI puts messages of its own into the conversation it holds with the model — verification gates when files have changed without anything being run against them, loop guards, close-out prompts. They are routine, they are not written by you, and usually you will not have seen them. Model providers reject two user turns in a row, so these have to be appended to your most recent message; without something marking the seam, the model receives our words as a continuation of your own sentence and answers you for something you never said.

They therefore arrive labelled, on a line of their own, in this shape: **AUTOMATED MESSAGE** · Cymela CLI · a1b2c3d4. That trailing value is a random token generated afresh each time the CLI starts. It exists only in memory for the life of that process and is never written to disk. The visible phrase is the part you can recognise on sight; the token is the part that makes it ours.

The token is there because a label a model is taught to obey would otherwise be a label anyone could type. The model is instructed that the marker carries authority only when the token is exactly right, and that the same marker bearing a wrong token, no token, or arriving inside a file, a fetched web page, a command's output, a commit message or a code comment is not from us and is not to be acted on. Text on your screen wearing that marker is not, by itself, evidence that Cymela wrote it.

The same rule applies in the other direction: a message you type wearing the marker is refused before it is sent rather than filtered afterwards. Nothing goes to your model provider, and the text you wrote is neither transmitted nor stored — it stays in the composer for you to edit. The refusal notice is shown briefly on screen and then disappears; it is not written to your session transcript, not recorded anywhere, and not reported to anyone — we receive nothing, because we operate no servers for the CLI to report to. Discussing the feature is not using it — the check looks for the marker dressed as a label, whether opening a line or emphasised in running text, which is the shape of an impersonation rather than a mention.

Its limits belong in this section with the rest. This guards the conversation against a message claiming an authority it does not have. It is not a security boundary, and it cannot make a model behave: whether the token check is honoured at all depends on the model you selected following the instruction to honour it, which — as Section 07 explains — varies enormously between models and is not something we control.

We give no assurance that these features are free of defects or that they will prevent any particular harmful outcome.

Third-party services

Using the CLI necessarily means using third-party services: a model provider, the npm registry, and — if you use the web-search tool — a search engine. These are operated by other people under their own terms and privacy policies.

You must comply with the terms of every third-party service you use through Cymela. Your agreement with a model provider is directly between you and that provider. We are not a party to it, we do not act as your agent in it, and we are not responsible for that provider's availability, pricing, output, data handling, or decision to suspend your account. If a provider terminates you for misuse, that is a matter between you and them.

Intellectual property

Ours

The Cymela name, logo, website design and written content are ours, and are protected by intellectual property law. Nothing in these Terms transfers ownership of them to you. The license in Section 05 is a license to use the software, not a grant of the brand.

Yours

We claim no ownership over your code, your prompts, or anything you create using the Services. Your work is yours. Because we never receive it, we could not claim rights in it even if we wanted to.

Models we publish

Any model weights we release are derived from a third-party base model and carry that base model's license with them, which is more restrictive than the license in Section 05. Our current research checkpoint, Hyper v1, is built with Qwen — it extends Qwen2.5-3B-Instruct and inherits the Qwen Research License: research and evaluation use only, not commercial use. That restriction is not ours to waive, and it travels to anything you build on the weights. The governing license text ships alongside the weights; read it before you download them, not after.

Published checkpoints are research artifacts, not products. They have received no safety training, they carry no warranty of any kind, and Sections 12 and 13 apply to them in full. Do not deploy them in front of users.

Generated output

Rights in model output are determined by the terms of the model provider that generated it and by the copyright law of your jurisdiction, which in several countries does not recognize machine-generated content as protectable at all. We make no claim to output, and equally give no warranty that output is original, non-infringing, or free of third-party rights. Verify before you ship, particularly for anything the model may have reproduced from its training data.

Disclaimer of warranties

The Services are provided "as is" and "as available", without warranty of any kind, whether express, implied or statutory. To the fullest extent permitted by applicable law, we disclaim all implied warranties, including any implied warranty of merchantability, fitness for a particular purpose, title, accuracy, and non-infringement.

We do not warrant that the Services will be uninterrupted, timely, secure, or error-free; that defects will be corrected; that any output will be accurate, complete, safe or reliable; or that the Services will meet your requirements. No advice or information obtained from us or through the Services creates any warranty not expressly stated here.

You use the Services at your own risk, and you are solely responsible for any damage to your computer system, loss of data, or loss of work that results.

Limitation of liability

To the fullest extent permitted by applicable law, we will not be liable for any indirect, incidental, special, consequential, exemplary or punitive damages, nor for any loss of profits, revenue, business, goodwill, data, or work product, arising out of or relating to your use of or inability to use the Services — whether the claim is framed in contract, tort (including negligence), strict liability or otherwise, and whether or not we were advised of the possibility of such damages.

This expressly includes loss or damage arising from: code the agent generated, modified, deleted or executed; output from a third-party model; the act or omission of any third-party provider; or another person's misuse of the Services.

To the fullest extent permitted by law, our total aggregate liability arising out of or relating to the Services will not exceed the greater of (a) the total amount you have paid us for the Services in the twelve months preceding the event giving rise to the claim, which for free Services is zero, or (b) EUR 100.

Indemnity

You agree to indemnify, defend and hold harmless Cymela and its operator from and against any claim, demand, action, loss, liability, damage, cost or expense — including reasonable legal fees — brought by a third party and arising out of or relating to:

• your use of the Services;
• your breach of these Terms, and in particular of Section 06 (Acceptable use);
• your violation of any law or of the rights of any third party, including intellectual property and data-protection rights; or
• content or code you generated, deployed or distributed using the Services.

This obligation does not apply to the extent the claim arises from our own wilful misconduct, and it applies only so far as the law of your jurisdiction permits.

Suspension and termination

You may stop using the Services at any time by uninstalling the CLI.

What termination means here

We would rather be straightforward about the mechanism than imply a power we do not have. The CLI runs on your computer, not on ours. There is no license server, no account system and no remote kill switch, and we cannot reach into your machine to disable a copy you have already installed. Termination is a legal act, not a technical one.

If you breach these Terms, the license granted in Section 05 ends automatically and immediately — without us having to do anything, or even to know about it. From that moment you have no right to use, copy or distribute the software, and continuing to do so is unlicensed use of a copyrighted work. That is an infringement in its own right, separate from whatever the misuse itself involved.

What we can do in practice

• Withdraw your license, as above, so that any further use is unlawful rather than merely prohibited.
• Block you from the repositories and platforms we control — including our GitHub and Hugging Face organizations — and refuse you support, issue responses, or future releases.
• Report the conduct to the model provider whose service you were using. Providers do operate accounts, and they can terminate yours. In most misuse scenarios this is the fastest and most effective consequence available, and it lies with them rather than with us.
• Report unlawful activity to the relevant law-enforcement or regulatory authorities.
• Suspend or terminate directly any hosted service, account or paid tier that may exist in the future.

We may take any of these steps at our discretion, with or without notice, where we reasonably believe you have breached these Terms or that your use exposes us or others to legal risk or harm.

Sections 06, 11, 12, 13, 14, 17 and 18 survive termination.

Changes to these terms

We may revise these Terms as the project develops. The version number and date at the top of this page will be updated when we do. Material changes will be flagged on the site rather than made silently. Continuing to use the Services after a revision takes effect means you accept the revised Terms; if you do not, stop using them.

Governing law and disputes

These Terms, and any dispute arising out of or in connection with them or the Services, are governed by the laws of the Republic of Serbia, without regard to its conflict-of-laws rules. The courts of the Republic of Serbia have exclusive jurisdiction, and you consent to that jurisdiction and venue.

If you use the Services as a consumer resident in the European Union, the United Kingdom or another jurisdiction whose law grants you a mandatory right to the protection of your local courts, this section does not deprive you of that right.

Before starting formal proceedings, please contact us at contact@cymela.com. Most issues can be resolved that way, and we would prefer to.

General

Severability. If any provision of these Terms is held unenforceable, it will be modified to the minimum extent necessary to make it enforceable, or severed if it cannot be, and the remainder stays in full effect.

No waiver. If we do not enforce a provision on one occasion, that is not a waiver of our right to enforce it later.

Assignment. You may not assign or transfer your rights under these Terms without our written consent. We may assign these Terms — for example, if the project is transferred into a legal entity formed to operate it.

Entire agreement. These Terms and the Privacy Policy, together with any separate license file distributed with a component, are the entire agreement between you and us regarding the Services, and supersede any earlier understanding.

No partnership. Nothing in these Terms creates a partnership, joint venture, employment or agency relationship between you and us.

Contact

Questions, notices and complaints about these Terms go to contact@cymela.com.

See also our Privacy Policy, which explains what data we handle — and, for the most part, do not handle.