Agent Security
Vyre Agent is designed with a defense-in-depth security model. You remain in full control of what Agent can access, modify, and execute at all times.
What Agent can access
By default, Agent can:
- Read files in any directory you've opened in Vyre.
- Write files within your current workspace.
- Run terminal commands in shells you explicitly open.
- Browse the web to read public URLs (docs, GitHub, Stack Overflow).
- Call MCP tools that you've explicitly configured.
Agent cannot:
- Access files outside your workspace without explicit permission.
- Access environment secrets unless they're in a file Agent can read.
- Make network requests on its own (web browsing is always visible in the tool log).
Permission model
Workspace-Scoped
Agent's file access is always scoped to the directories you've opened. It cannot traverse outside your workspace root.
Explicit Tool Approval
In strict mode, Agent asks for explicit approval before running any terminal command. Review the command before it executes.
Full Audit Log
Every tool call Agent makes — every file read, file write, and command run — is recorded in the session log. Nothing is hidden.
MCP Allowlist
Only MCP servers you explicitly configure in Settings are available to Agent. Connections require manual setup.
Privacy controls
Code privacy
Your code is sent to the AI model to generate responses. Vyre uses the following protections:
- Code is transmitted over TLS 1.3.
- Vyre does not train its own models on your code.
- For Anthropic/OpenAI models, Vyre uses API contracts that prohibit training on API inputs.
- Vyre Business and Enterprise plans include a Zero Data Retention option where code is never logged by the model provider.
Gitignore respect
Agent respects your .gitignore by default. Files that are gitignored (e.g., .env, secrets.json) are excluded from context unless you explicitly attach them with @filename.
You can also create a .vyrreignore file to add additional exclusions specific to Vyre.
Local model support
For maximum privacy, Vyre supports local inference via Ollama. Configure a local model in Settings → Models → Add Local Model. With a local model, no code ever leaves your machine.
Enterprise controls
The following controls are available on Vyre Enterprise plans.
Enterprise administrators can configure:
- Model allowlisting — restrict which AI models agents can use.
- Tool restrictions — disable specific tools (e.g., disable web browsing for regulated industries).
- Approval gates — require a second team member to approve large agent runs before execution.
- Data residency — choose where data is processed (US, EU, or self-hosted).
- SSO enforcement — require all team members to authenticate through your identity provider.
- Audit log export — export all agent activity logs to your SIEM or compliance tooling.