Vetta centers work in a local workspace, but “runs locally” does not mean that all content stays on the device. Model providers, MCP, and plugins may send task data to configured external services. Understand each boundary before using it.
Data flow overview
Local tools may continue to read or modify files. MCP, plugins, webhooks, and IM channels can create separate external requests.
Responsibility by boundary
| Boundary | Data it may see | Your check |
|---|---|---|
| Model provider | Prompt, citations, history, tool results, and images | Provider, base URL, model, and privacy terms |
| Local tools | Workspace files, command output, and processes | Sandbox or full access, target path, and command |
| MCP | Tool arguments and external results, OAuth or token metadata | Source, transport, and auto-approval scope |
| Plugin | Manifest-declared UI, file, network, or Agent capabilities | Source, permissions, version, and enabled state |
| Knowledge base | Imported sources, processed results, and retrieved excerpts | Processing model, import scope, and deletion semantics |
| Webhook / IM | Notification content, conversation, or attachment metadata | Destination, channel credentials, and sharing scope |
| Paired phone | Session text, attachments, the desktop picture, and pointer or keyboard input | Give the QR code, connection code, and password only to the intended phone, and remove phones you no longer use. The relay forwards ciphertext only |
Credential handling
- Enter credentials only in dedicated model, MCP, IM, or webhook configuration UI.
- Do not put keys, tokens, cookies, or authorization codes in tasks, project files, skills, logs, or issue reports.
- References such as
env:MY_API_KEYstore only the variable name; the environment that starts Vetta must still protect the value. - Inspect diagnostic exports and sessions before sharing them. Do not assume every third-party tool output is automatically redacted.
Execution permissions
Restricted sandbox mode limits file and process access to the workspace capability boundary. Full access permits broader system operations. If a platform cannot provide sandboxing, Vetta explains why; do not treat “sandbox unavailable” as equivalent protection.
A permission prompt only means that the host is asking whether to perform an action. Confirm the target path, command, network address, and reversibility, and do not assume the request itself is trustworthy.
Installing external capabilities
Skills, plugins, MCP servers, and external capability repositories are untrusted inputs. Install only from trusted sources, inspect updates, and grant the smallest permission set that meets the goal. Disable to keep an installation; remove to uninstall it.