# OneAgent for VS Code privacy policy

Effective date: 2026-08-11

Source update: 2026-10-24, common skills and optional local memory candidate;
not evidence of Marketplace publication.

OneAgent for VS Code runs an agent on the user's computer through a local
OneAgent Host and a tenant-bound OneAgent Gateway.

## Data processed to provide a turn

When a user submits a request, the prompt and any explicitly attached context
may be sent through the tenant-bound OneAgent Gateway to the tenant's configured
model provider. Files and command output are read locally by the agent when
needed for the requested work. OneAgent does not add prompt text, file contents,
image contents, or command output to its audit or reliability telemetry.

The configured model provider processes request content to generate the agent
response. Provider terms and retention controls associated with the tenant's
provider credential also apply.

## Data stored locally

Threads, the Agent Home, caches, tenant profile artifacts, and change snapshots
remain on the user's device. OAuth refresh credentials and device secrets are
stored with VS Code SecretStorage. Passwords and provider API keys are not
stored by the extension.

Optional OneAgent memory stores explicit notes and preferences in a SQLite
database under the authenticated local profile, separate from skill assets.
Workspace notes are workspace-scoped; UI-saved profile preferences can be
retrieved across that profile's workspaces. Relevant memory may be included in
subsequent model requests. **OneAgent: Manage Memory** supports viewing,
correction, forgetting, export, pause and clearing. Automatic conversation
summaries are off by default. When enabled, bounded finished-turn text is sent
through the existing authenticated Gateway/model allowance, not a separate
Claude-Mem service. Summary retention is configurable; common secret patterns
are redacted, but users must not save secrets. Thread deletion removes associated
notes and queued captures; independent explicit profile preferences remain until
forgotten or cleared. Clearing memory keeps conversation history unchanged.

Preparing common tools is explicit and downloads dependencies from the named
upstream registries. Some optional OCR/layout and semantic-search operations
download model assets on first use. Tool caches stay in the profile and local
indexing is limited to requested workspace paths; no whole-home indexing is
started by extension activation.

Uninstalling the extension preserves local Agent Home data by default so an
accidental uninstall does not destroy work. The user can run the OneAgent purge
command to remove the selected tenant's local profile after reviewing the
confirmation prompt.

## Service metadata

OneAgent retains content-free audit metadata for up to 90 days. This may include
tenant, user, device, model, request identifier, timestamps, status, latency,
and usage counters. It does not include prompts, file contents, command output,
or response text.

Reliability telemetry is off by default during the pre-release. If enabled in
a later release, it will be consented, content-free, documented here, and
separable from tenant business data.

## Tenant isolation

Tenant membership, provider policy, quota, and audit metadata are isolated by
the authenticated tenant identity. Local Agent Homes use opaque tenant/user
profile keys and are not shared between memberships.

## Contact

Privacy questions may be sent to `charles@one-agent.ai`.