OpenClaw Privacy: What Data Your Agents Actually Send and Store

Running an AI agent on sensitive data without understanding the data flow is the mistake I've seen cause the most compliance problems. You think the data stays local. It doesn't — at least not all of it, not by default.

What Actually Leaves Your Machine

OpenClaw is a runtime that orchestrates AI model calls and tool use. Every time your agent processes a message, that conversation context — including any file contents, tool results, or prior messages — gets packaged into a request and sent to your model provider.

That is the primary data flow. Everything else is secondary.

Here's where most people get caught out: they focus on "does OpenClaw collect data?" and miss the more important question — "does my AI provider retain data?" OpenClaw doesn't collect anything. Your AI provider might. Those are different questions with different answers.

Model Provider Data Flow

The data your agent sends to its configured provider depends on the provider's API and your session configuration. Here's the breakdown for the three most common setups:

Both Anthropic and OpenAI explicitly state that API traffic — as distinct from consumer product traffic — is not used for training by default as of early 2025. Verify this against each provider's current data processing agreement before deploying.

Local Storage Map

OpenClaw writes to your local file system in several locations. Here's what gets stored where by default:

• ~/.openclaw/config.yaml — your main config file, including API keys (if stored directly rather than via env vars). Keep this protected.
• ~/.openclaw/sessions/ — session logs, only created if logging.enabled: true in config. Off by default.
• ~/.openclaw/skills/ — installed skill definitions. No sensitive data unless skills themselves write sensitive output.
• ~/.openclaw/cache/ — model response cache if enabled. Can contain conversation content if caching is on.

None of these directories sync anywhere automatically. OpenClaw does not have a sync or backup feature. If you've set up your own file sync (Dropbox, iCloud, OneDrive), make sure these paths are excluded — especially config.yaml, which contains your API keys.

Tool Use and Third-Party Services

Every tool your agent uses is a potential outbound data channel. This is the area most people overlook entirely.

Sound familiar? You set up web search, calendar integration, and a database connector, and now you're not sure exactly what those tools are transmitting. Let's go through the common ones.

Web Search Tools

When your agent performs a web search, the search query — which may contain context from your conversation — is sent to the search provider. Brave Search, DuckDuckGo, and Google all receive the query string. Review what context your agent includes in search queries before connecting sensitive data sources.

File System Tool

This is the big one. When your agent calls the file system tool, the file contents are read into the agent's context and sent to the model provider with the next inference request. A 50KB document becomes part of the inference payload. The model provider receives and processes that content.

Calendar and Productivity Integrations

Calendar skills (Google Calendar, Apple Calendar) use OAuth and read/write event data through the provider's API. Event titles, descriptions, attendees, and times all pass through the skill to the model context. That context then goes to the AI provider.

Running OpenClaw Fully Air-Gapped

If zero external data transmission is a requirement, this configuration works as of early 2025:

# config.yaml — air-gapped configuration
providers:
  default: ollama
  ollama:
    base_url: "http://localhost:11434"
    model: "llama3.2:latest"

logging:
  enabled: false

tools:
  web_search: false
  external_apis: false

# Do not configure any external integrations

With this config, all inference runs on localhost via Ollama. No data leaves the machine. The tradeoff is model capability — local models are less capable than frontier API models for complex reasoning tasks.

We'll get to the specific model performance tradeoffs in a moment — but first, note that "air-gapped" still means your conversations and file contents are processed by Ollama locally. If you're running Ollama on a shared server, that server operator can access the data.

Common Privacy Mistakes in OpenClaw Deployments

1. Storing API keys directly in config.yaml without permissions hardening. The file defaults to world-readable on some systems. Use chmod 600 ~/.openclaw/config.yaml immediately after setup.
2. Enabling conversation logging and forgetting it's on. Session logs accumulate indefinitely and can contain sensitive file contents, credentials your agent encountered, and private user messages.
3. Not reviewing third-party skills before installing. Skills run in the same process and have access to the same data your agent does. Treat skill installation like installing any npm package.
4. Assuming "no telemetry" means "no external calls". OpenClaw has no telemetry. But your configured tools, skills, and the model provider all make external calls. Map every outbound connection before handling sensitive data.
5. Using the same OpenClaw instance for personal and work data. Session history and context can bleed across conversations if you don't clear sessions between sensitive use cases.

Frequently Asked Questions

Does OpenClaw send my conversations to Anthropic or OpenAI?

Every message you send through OpenClaw goes to whichever AI provider you've configured. The provider receives the full conversation context. OpenClaw itself does not retain copies unless you've enabled conversation logging in your config.

What data does OpenClaw store locally?

By default: config files, skill definitions, and optionally session logs if enabled. Storage is at ~/.openclaw/ on Linux/macOS. Nothing is sent to any third-party analytics or telemetry service.

Can OpenClaw read my private files without permission?

OpenClaw only accesses files that tools explicitly request. Configure allowed_paths in your config to restrict which directories agents can access, creating a hard boundary at the file system level.

Does OpenClaw send telemetry data to its developers?

No. OpenClaw's default builds include no telemetry. You can verify this by reviewing the source code on GitHub — there are no analytics calls or usage trackers in the core runtime.

How do I run OpenClaw with no external data transmission at all?

Use a local model via Ollama as your provider. With a local model, no conversation data leaves your machine. Disable web search and external tool integrations in your config for a fully air-gapped setup.

What happens to my data when I use a third-party skill from ClaWHub?

Third-party skills run in the same process as OpenClaw and can access anything your agent can access. Review the skill's source code before installing untrusted skills — treat them like any open-source dependency.

Does OpenClaw comply with GDPR or other data regulations?

OpenClaw is a self-hosted tool — GDPR compliance is your responsibility as the operator. Your obligations depend on what data your agents process and where you run them. Review your AI provider's DPA separately.

You now have a complete map of every data flow in an OpenClaw deployment. The model provider receives your conversation context — that's the primary flow to manage. OpenClaw itself doesn't collect anything. Tools and skills are the secondary channels to audit.

Start by reviewing your provider's data processing agreement and locking down your config file permissions. Both steps take under five minutes and close the two most common privacy gaps we see in OpenClaw deployments.