For AI agents: The documentation index is at https://docs.digitalocean.com/llms.txt. Markdown versions of pages use the same URL with index.html.md in place of the HTML page (for example, append index.html.md to the directory path instead of opening the HTML document).

We do not store model inputs or outputs on DigitalOcean infrastructure, except for conversation data that DigitalOcean Signals collects after you enable it. For details, see DigitalOcean Signals.

For more information about DigitalOcean’s security practices, see our security page.

DigitalOcean Hosted Models

Inference requests for DigitalOcean-hosted models run entirely within DigitalOcean’s infrastructure. Input is stored in the local browser and sent to the model for inference on DigitalOcean’s infrastructure. The returned output is then stored in the local browser’s storage and displayed in the agent’s interface. If you’ve configured your agent to use prior parts of the conversation as additional context for output, the agent accesses the browser storage as necessary to retrieve the context.

Customer data submitted to the model is never sent back to the original model creator, so the creator cannot use it for training. This is different from using a model through its creator’s own API, such as querying a DeepSeek model directly through api.deepseek.com, where that provider’s data retention and training policies apply. DigitalOcean does not use this data to train, retrain, or fine-tune any model, and does not share it with any third parties for training or fine-tuning.

For custom interfaces and applications you have developed to use DigitalOcean Inference, you choose where to store this data.

Third-Party Models

We do not store agent input or output when using third-party model providers.

• OpenAI models: When using OpenAI models on DigitalOcean Inference, OpenAI’s zero data retention policy excludes customer content from abuse monitoring logs, treats the store parameter for /v1/responses and /v1/chat/completions as false, and may still allow certain endpoints to store limited application state.

• Anthropic models: When using Anthropic models on DigitalOcean Inference, Anthropic applies a Zero Retention Policy and deletes prompts and outputs after generation, except where required by law or to address malicious use. Anthropic may run risk classification models on prompts and outputs to detect potential Acceptable Use Policy or agreement violations.

  For Claude Fable 5.1 and Claude Fable 5, access requires a mandatory 30-day data retention of prompts and completions for trust and safety reviews. You do not need to configure workspace routing. DigitalOcean Inference automatically handles routing upstream to a compliant workspace.

  The Files API (/v1/files) uses a separate retention model. Files ingested via this API, including large documents used as intermediate context processing states, are stored for reuse across multiple API requests and are not automatically purged. They are retained statefully on the upstream infrastructure until an explicit, authenticated DELETE request is executed via the API. Because of this persistent storage behavior, the Files API pipeline does not qualify for Zero Data Retention (ZDR) frameworks and is not HIPAA-compliant.

Data sent to other third-party model providers is handled according to the applicable provider’s policies.

Model Synthesis Tool

Model synthesis tool requests are not stored by DigitalOcean. Retention follows the same policies as the underlying model calls. A model synthesis tool request is evaluated by every model on the panel, so model selection is a data-governance decision. Confirm that your prompt and context are appropriate for each provider in the panel. For provider terms, see Third-Party AI Models. For how to configure and call the tool, see How to Use the Model Synthesis Tool.

Batch Inferencing

We are a pass-through for OpenAI and Anthropic models and do not retain your batch inputs or outputs. OpenAI and Anthropic storage policies apply.

After the retention window, all input and output artifacts are permanently purged and cannot be recovered. Download your results before the expiration window closes.

DigitalOcean Signals (Public Preview)

Key definitions

• Trace: A record of an agent or inference interaction, which may include End User prompts and inputs, agent calls to tools and databases and their parameters, retrieved data, intermediate steps, model responses, and associated metadata such as timestamps, identifiers, latency, cost, and error states.
• Session: A set of Traces grouped as a single End User interaction or conversation.
• Signal: A derived attribute, label, score, or classification produced by applying logic or models to Traces or Sessions.
• End User: An individual or system whose interactions with your agent generate Traces. End Users are your users, not DigitalOcean’s.

Data we process

When you consent to Signals for Harness Runtime agent or for Serverless Inference, or use it with Simulations or Evaluations, we process the Trace content generated, including the text of End User prompts and inputs, tool and database calls and their parameters, retrieved content, model responses, and reasoning or intermediate steps; the Session groupings and metadata associated with those Traces; and the Signals we derive from them. We also process Usage Data about your use of the feature.

A note about personal data. Trace content is processed in raw form and may contain personal data, or special-category or sensitive data, if that data is present in the underlying agent or inference interactions. You determine what your models, agents, and End Users submit.

How we use data

We use Trace content and Signals only to provide Signals to you and support it: to group Traces into Sessions and annotate Signals; to make Sessions and Signals available to you through dashboards, filtering, and search; to enable you to reuse qualifying Traces for your own purposes, such as with Evaluations and Simulations; to detect behavioral and operational issues in your agentic and inference workflows; and to operate, secure, troubleshoot, and improve the feature for you. We may use anonymized, aggregated Usage Data to improve the Service.

We do not use your content to train generalized AI models

DigitalOcean does not use your Trace content, Signals, inference inputs and outputs, agent inputs, or agent outputs to train, fine-tune, or improve any generalized artificial-intelligence or machine-learning model, including any foundation model offered to other customers. We process that data only to provide Signals to you or as required by law.

Signals may capture responses from the models your agents call. Those models are operated under their own data-use and retention practices. For models made available through DigitalOcean’s inference infrastructure, provider-specific data-use and retention details, including any zero-data-retention arrangements or mandatory retention periods, are described in the DigitalOcean Serverless Inference data privacy documentation. For Third-Party Models you connect directly, those practices are governed by your relationship with that provider, not by DigitalOcean.

No cross-customer use

We do not combine your Trace content with other customers’ data, and we do not use your Trace content or Signals to develop products for, or to benefit, other customers. Signals are scoped to your account.

Automated redaction of sensitive data

DigitalOcean applies automated redaction that is designed to detect and replace common categories of sensitive data before content is shown to you in the Console. The categories we aim to detect include personal identifiers such as names, contact details, locations, and dates; government and national identifiers and documents; financial data such as payment card and bank account numbers; medical and clinical terms; network identifiers; and credentials and secrets such as API keys, access tokens, passwords, private keys, and database credentials. Detected items are replaced with typed placeholders.

Automated redaction is a safeguard, not a guarantee. It is not infallible and may not detect every item, particularly in languages or formats it does not fully support. It does not change your responsibilities for the data your models, agents, and End Users submit, and it does not by itself satisfy your obligations under HIPAA, PCI-DSS, or other industry-specific regulations. You remain responsible for determining whether Signals is appropriate for your data and regulatory environment.

Your content, your control

As between you and DigitalOcean, you retain all rights in your Trace content. For Harness Runtime, you can enable or disable Signals through the DigitalOcean Control Panel, CLI, or API. For Serverless Inference, you can enable or disable Signals through the API. Disabling Signals stops further Trace capture for that agent or model. You may request deletion of collected data through the Signals data deletion API.

For personal data contained in Trace content, you are the Controller and determine the purposes and means of processing your End Users’ data, and DigitalOcean is the Processor, processing Trace data to deliver Signals and on your instruction to enable the feature data.

Data residency and regions

Trace content and Signals are processed and stored in a DigitalOcean United States data center region. Signals does not offer regional data storage or data localization at this time.

Traces collected for Signals are retained in accordance with DigitalOcean’s standard retention practices and the applicable documentation. You may request deletion of collected data through the Signals data deletion API. For how the models your agents call process and retain the data you route to them, see the DigitalOcean Serverless Inference data privacy documentation.

Data minimization

Signals is designed to surface the most informative Sessions rather than store all raw traffic, which reduces the volume of Trace data retained.