Privacy without absolute promises
A private AI workspace with explicit boundaries
LocalAI Cowork is local-first: the application and workspace data live on your computer. Privacy still depends on the model, tools, and services you choose for each task.
Data that stays in the desktop workspace
Projects, tasks, settings, audit records, and other application state are stored locally in the app data directory. Provider keys and connector secrets use the operating-system credential vault rather than ordinary frontend settings. Sensitive fields are removed from persisted frontend state and bounded log output is redacted.
When task data leaves the device
A remote model cannot answer without receiving a request. If you select OpenRouter or another remote OpenAI-compatible endpoint, that provider receives the prompt and context sent for the task. A remote MCP server, connector, web tool, or terminal command can likewise transfer or expose information according to its behavior. LocalAI Cowork does not claim those third parties are private.
| Configuration | Expected boundary |
|---|---|
| Ollama on the same device | Model inference stays local, subject to Ollama and model configuration. |
| Hosted model endpoint | Prompt, selected context, and tool-related messages can reach the provider. |
| Local MCP server | The process runs locally but can still have its own network access and data handling. |
| Remote MCP or web tool | Information required for the requested action is transferred to that service. |
| Website | No analytics, ad scripts, third-party fonts, or tracking cookies are shipped. |
A safer evaluation workflow
- Begin with an Ollama profile and a copy of non-sensitive files.
- Grant the smallest folder and tool scope that can complete the task.
- Keep destructive or external actions behind approvals.
- Inspect the output and audit trail; do not assume model-generated work is correct.
- Review every provider and MCP server separately before giving it real data.
- Report suspected vulnerabilities through GitHub private vulnerability reporting.
No hidden account or website tracking
The app does not require a LocalAI Cowork account, and the current release has no implemented telemetry sender. The project website is static. It does not load analytics, ads, trackers, remote fonts, or cookie consent code. GitHub, download hosts, and any external link have their own policies.
Inspect the claim
The repository contains the credential store, persistence rules, redaction tests, privacy documentation, security policy, dependency scans, SBOM generation, and release workflow. Open source makes the boundary reviewable, while tests and reports provide evidence that can be improved through public contributions.