If you run your own model, you still need to know what happens to anything you send to an external service. A Tabstack Research call sends your query to our hosted API. Contracted model providers process the query and web content to produce the answer.
We describe Tabstack as “private by default.” This post explains what we retain after a request and what the detailed data collection setting changes.
What happens during a Research call
Suppose your query is “What changed in the latest Node.js LTS release?” Tabstack receives that query and reads public sources to produce the answer. Your application can then pass the returned report to your own model.
flowchart TB A[Your application] -->|Query| B[Tabstack hosted API] B -->|Report and cited pages| A B --> W[Public web, through a shared page cache] B --> P[Contracted model provider] B -.-> R[Retained: request metadata always, payloads for 90 days while detailed data collection is on]
Your application calls the API with your key over TLS 1.2 or higher. We fetch the public pages the answer needs, and our direct fetches identify themselves as Mozilla-Tabstack/1.0 (+https://tabstack.ai). Your query and the page content go to a contracted model provider, which produces the report. The report and the pages it cites come back to your application. All of this runs on our hosted infrastructure, not inside your local model.
What we retain after the request
Request metadata is recorded on every request, on every plan: the endpoint, success or failure, credits spent, timestamps, and the organization and key that made the call. The Data Handling page doesn’t state how long request metadata is kept.
Payloads are the target URL, the request parameters, the response data and the output. For Research, that includes your query. We store payloads only while your organization has detailed data collection on. New organizations start with it on, and an organization admin can turn it off.
With detailed data collection on, payloads are kept for 90 days and appear in your console, which helps when you’re debugging results or auditing what a task did. With it off, new payloads aren’t stored. If your workload can’t have payloads stored at all, turn the setting off before you send anything.
Fetched page content is cached by URL, effort and region, with a short time-to-live. The cache isn’t per account, so public page content fetched for one account can be served to another. That matters when you need a fresh copy of a page. Pass nocache: true to bypass the cache for that request.
Our Privacy Notice is the canonical statement of what we retain. It applies to the API as well as the website and our other services, and it says: “We store history (including your content and/or Outputs) for 90 days or unless you delete it.” The notice also asks you to be careful about putting confidential or sensitive information in your inputs. For privacy questions, our Data Protection Officer is at dpo@mozilla.com.
What “never trained on” covers
Our Data Handling page states it directly: your requests and the content Tabstack processes for you are not used to train models.
Research is one of the endpoints that sends your query and the page content to a contracted model provider. The Privacy Notice names Google Cloud, including Vertex, and OpenAI as contracted providers. Each provider’s own retention and training terms for that traffic are a separate question, and the public documents don’t list them per endpoint.
When a hosted API is not suitable
If your workload requires that data stays inside an environment you control, a hosted Research call is the wrong tool, and no setting changes that. There is no self-hosted deployment of the API. Pilo, our open-source automation engine, runs browser automation in your own environment. It covers automation only, not Research.
Where to review the details
Tabstack is built by Mozilla, which is why we state these limits as plainly as the commitments. Read our Data Handling documentation for the request flow and retention settings. If your review requires provider-specific terms, contact us before sending sensitive data.
