Privacy Policy

1. Who is responsible for your data

AgentOp is the data controller for the personal data described here. For any privacy question, or to exercise the rights in section 8, contact the operator's public community channel .

2. What the standard runtime keeps on your device

During ordinary use, the standard runtime handles the following locally:

• With a local model, the prompts you type and the answers it gives. Cloud providers receive the prompts needed to answer.
• Files you open with an agent — documents, PDFs, images, audio. Reading happens in your browser. Where an agent indexes a document for search, the index is stored in your own browser's storage (OPFS or IndexedDB) and is not uploaded to AgentOp by the standard runtime. Relevant excerpts can be sent to a cloud model when you select a cloud provider.
• Your OpenAI or Anthropic API key in readable form. It is encrypted client-side with AES-256-GCM before it is stored or embedded in a downloaded agent.
• The contents of your microphone recordings or the transcripts made from them — speech-to-text runs on your device.
• Any detailed profile of your hardware. The pre-flight browser check measures your machine in your browser, keeps the full report there, and transmits only a coarse bucket (section 3).

We also do not sell personal data, do not share it with data brokers, and do not use your content to train any model.

3. What we do store

Your account

Username, email address, a hash of your password (never the password itself), and whether your address is verified. Optionally, anything you choose to add to your profile: bio, website, GitHub and X usernames, company, job title, location, and your theme and visibility preferences. We also store the Terms version accepted at signup and the time of acceptance. Existing accounts without a recorded acceptance are not automatically marked as having agreed.
Lawful basis: performance of our contract with you.

What you build

The agents and templates you create — their name, description, Python code, prompts, package list, tags, and revision history — plus ratings, collections, follows and notifications. Private agents are not publicly listed or accessible to other ordinary users; their source is still held in our database. Note that an agent's code and prompts are stored (we have to store them to generate the file). Ordinary conversations are not saved by that generation process.
Lawful basis: performance of our contract with you.

Where your account came from

One row per signup recording the referring host — the domain only, never the full URL, which can contain a search query or a private thread — any UTM tags on your first visit, the page you landed on, and, if you answer the one-tap question after signing up, how you say you heard about us. Nothing is written into the session for an ordinary visit with no UTM tags and no external referrer. Each completed signup still creates an attribution row, which may initially have empty source fields.
Lawful basis: legitimate interest in understanding which channels reach people, balanced by keeping the record to a channel rather than a journey.

Usage milestones on agent pages

When you run an agent on this site, the page reports milestones — the agent loaded, a model finished downloading, a query completed — so we can tell "nobody visited" apart from "it failed on their machine". It counts events; it never reads message contents. The platform's funnel tracker applies to the hosted run page. Downloaded agents do not include it, but can still contact model hosts, runtime-asset hosts, cloud providers and our model registry.
Lawful basis: legitimate interest in knowing whether the product works.

A coarse hardware bucket

The browser check classifies your machine locally and automatically sends a coarse result when its check completes. You can also choose to attach it to feedback. We receive only a summary: a performance tier, one GPU vendor word, which runtime build applies, and two yes/no flags. The detailed measurements stay in your browser, because together they would be an identifying fingerprint.
Lawful basis: legitimate interest in diagnosing device compatibility and knowing whether the product works, balanced by sending a coarse bucket rather than the measurements themselves. It is sent automatically, so it does not rely on your consent; you can object using the contact route in section 1.

Feedback you send

The category, your message, the page URL, your account association and the hardware bucket above if you attached it. A report is stored and also emailed to the configured feedback inbox. Include only information you intend to share.
Lawful basis: consent, and our legitimate interest in fixing what you reported.

Saved test cases and evaluation results

When a creator uses the test-suite feature, we store test inputs, expected results, assertions and submitted agent outputs, alongside the agent, template, provider, account and run time. The evaluation panel sends outputs back for scoring and history. Only put information in test cases that you intend to store here. These records remain until removed with their associated agent, template or account; test cases can also be deleted individually.

Server logs

Standard operational logs: the requested path, method, response status, timing, and the username if you were signed in. Used for debugging and abuse investigation.
Lawful basis: legitimate interest in operating and securing the service.

4. Sponsored AI generation

The optional sponsored AI generator creates a template using our own API key, at our cost. Because we pay for it and have to prevent abuse, the generation request is handled by our infrastructure. For each attempt we store: your account, the calendar date, your IP address, which model was used, prompt and completion token counts, an estimated cost, a success or error code, and any validation warnings. Your description is sent to OpenAI to fulfil the request under their API terms.

The generated template and agent are also saved to your account. This audit record does not apply to ordinary bring-your-own-key inference or local inference. If you never use the sponsored generator, no sponsored audit row is created for you.
Lawful basis: performance of our contract, and legitimate interest in preventing abuse of a feature we fund.

5. Cookies and analytics

Essential cookies. A session cookie to keep you signed in and a CSRF token to protect forms. These cannot be turned off without breaking the site.

Google Analytics 4. We use GA4 for aggregate traffic statistics — pages viewed, rough geography, device category, and the funnel milestones in section 3. The site loads Google Tag Manager and GA4 on site pages; they can set cookies and receive page and device metadata. Our funnel events are designed to exclude prompts, agent output, file contents and API keys. Third-party scripts can nevertheless access page contents. See Google's privacy policy. You can block these scripts with a standard content blocker or browser tracking protection and the site will work normally.

Your browser's own storage. Agents keep things locally that never leave your device: downloaded model files, an encrypted API key if you saved one, a document index if you built one, and small preferences such as your theme or whether text-to-speech is on. The browser check page manages cached model files, not all local data. Use your browser's site-data settings to clear other storage; that can remove saved keys, documents and preferences and sign you out.

6. Who else is involved

Services involved in delivering the site include:

• DigitalOcean — application hosting and the managed database (data stored in the EU or US region as configured).
• Amazon SES — transactional email (verification, password reset, notifications).
• Google Analytics — aggregate traffic statistics, as described above.
• OpenAI — sponsored generations, as described in section 4; OpenAI or Anthropic also receive inference requests directly from your browser when you select them.
• Cloudflare Turnstile — anti-abuse checks for sponsored generation, when configured.
• Sentry — error reports, when enabled, containing stack traces and request metadata.

Model weights for local agents are downloaded by your browser directly from third-party hosts such as Hugging Face. That request comes from you, not from us; their privacy policy applies to it and we never see it.

Processing locations and transfer arrangements depend on the provider and deployment configuration. Contact us for the applicable arrangements.

7. How long we keep things

Deleting your account removes your published content from the service. Copies other users downloaded or forked before then are outside our control — a downloaded agent is a file on somebody else's computer.

8. Your rights

Depending on where you live, you have the right to access the personal data we hold about you, to correct it, to delete it, to restrict or object to certain processing, to receive a portable copy, and to withdraw consent where processing relies on it. You can also complain to your local data protection authority.

You can edit your profile and manage your agents and templates in the site. There is currently no self-service account-deletion or personal-data-export screen. To request account deletion, a data export or other assistance, contact the operator's public community channel and we will respond within 30 days.

9. Children

The service is not intended for anyone under 16. We do not knowingly collect personal data from children. If you believe a child has created an account, contact us and we will remove it.

10. Security

Traffic is served over HTTPS with HSTS. Passwords are hashed, not stored. API keys are encrypted in your browser with AES-256-GCM using a key derived by PBKDF2 at 600,000 iterations, so what reaches storage is ciphertext we cannot decrypt. Generated agent files are built on demand and never written to disk on our servers — there is no folder of generated agents to leak.

No system is perfectly secure. If we become aware of a breach affecting your personal data, we will notify you and the relevant authority as required by law. If you find a vulnerability, please report it to the operator's public community channel .

11. Changes to this policy

The version and date at the top of this page identify the current document. If we make a material change to what we collect or why, we will notify account holders before it takes effect.