Security and privacy

switchboard-ai-sdk is a TypeScript SDK for Node.js and Electron apps that lets developers discover and use local AI tools installed on a user's machine, including Codex, Claude Code, OpenCode, and Ollama, through one consistent API.

This page explains the security boundaries.

What runs locally

• The switchboard-ai-sdk code runs inside your Node.js or Electron process
• Discovery scans the local machine for installed CLI tools and the Ollama HTTP server
• Chat requests are sent to the local provider, not to a switchboard-hosted service

What data leaves the machine

The SDK itself does not send data to remote servers. However, the underlying provider may:

Provider	Data destination
Ollama	Local inference; no network call unless the user configured a remote Ollama host
Codex	OpenAI's services when the CLI makes cloud model calls
Claude Code	Anthropic's services when the CLI makes cloud model calls
OpenCode	OpenCode-hosted or OpenAI-backed models depending on configuration

Review each provider's privacy policy before processing sensitive data.

What depends on the provider

• Auth state is managed by the provider CLI
• Model availability is determined by the provider
• Sandbox and file-system access are controlled by the provider
• Error messages and output formats come from the provider

Sandbox limitations

Agent tools can read and sometimes write files:

• Codex offers sandbox modes: read-only, workspace-write, and danger-full-access
• Claude Code and OpenCode operate according to their own CLI behavior

Always choose the most restrictive sandbox level that still lets the tool do its job.

Electron security recommendations

• Run switchboard-ai-sdk in the main process, not the renderer
• Use contextIsolation: true and nodeIntegration: false
• Expose only the minimum surface through a preload script
• Bind the HTTP bridge to 127.0.0.1 and never expose it to the network
• Validate all messages from the renderer before acting on them

Example secure preload:

contextBridge.exposeInMainWorld("switchboard", {
  getUrl: () => ipcRenderer.invoke("switchboard:url"),
});

Why browser-only apps should not expose arbitrary local AI access

Browser pages can load untrusted third-party scripts. If a browser page could directly call local AI tools, malicious code could:

• read local files through agent tools
• run expensive or harmful prompts
• exfiltrate data through provider calls

Always route browser access through a trusted local backend process such as Electron's main process or a Node.js server.

Best practices

1. Inform users before connecting to agent-capable tools
2. Default to the least permissive sandbox mode
3. Check tool.type and tool.capabilities before use
4. Never ship API keys for hosted providers inside the SDK; rely on the user's local auth
5. Keep the optional HTTP bridge on localhost only

Reporting security issues

If you find a security issue in switchboard-ai-sdk itself, please open a private issue on GitHub.