# Security & permissions

<a id="what-pane-does-not-do"></a>

Pane runs your coding agents in terminals on your own machine. Your code leaves the machine only when an agent you run, like Claude Code or Codex, sends context to its provider. Pane doesn't proxy those requests, and it doesn't sandbox the agent (see [Agent permissions](#agent-permissions)).

## Local-first by default

There's no cloud sync. Your repos and [worktrees](/docs/panes-and-worktrees) stay where you put them on disk, and opening a pane uploads nothing. For local coding, Pane checks GitHub for updates, fetches model prices from OpenRouter for cost estimates, and sends the product analytics described below. Optional features, such as voice dictation, the leaderboard and remote access, also use the network.

## Phone access and Pane's relays

When you [pair a phone with a QR code](/docs/phone), your phone reaches your computer directly when it can, and through one of Pane's relays when it can't. The encryption (standard TLS 1.3) ends on your computer, so the relays can't read what passes through. They see which computer a connection is for, IP addresses, timing and byte counts. The plain-words version, including what a security review checked, is [Security, in plain words](/docs/remote-security). You can also [run your own relay](/docs/self-hosted-relay).

Tailscale and `pane-remote://...` connection codes still work, and they don't use Pane's relays. See [Tailscale and Remote VMs](/docs/remote-daemon).

Remote access is free and open source, with no subscription. You provide the machine and any provider credentials your agents need.

## Open source, AGPL-3.0

Pane's source is at [github.com/greenfield-inc/Pane](https://github.com/greenfield-inc/Pane). You can audit it, fork it, or build on it. AGPL-3.0 means network-served derivatives stay open source under the same license.

## Telemetry

Product analytics are on by default. Pane uses [PostHog](https://posthog.com/privacy) to record which features you use (pane creation, agent launch, keyboard shortcuts) and your OS and app version. If it finds them on your machine, Pane ties those events to your GitHub username, or your git email or name, from `gh` or your global git config. So these analytics aren't anonymous.

Pane never sends:

- File contents or code from your repos
- Agent prompts or responses
- API keys or secrets
- File paths beyond worktree names

The [`runpane` CLI](/docs/runpane-cli) sends no events for ordinary commands that succeed. It reports failures, as an error code and category, and its installs and updates: when a download is requested, when it succeeds, and when it falls back to downloading from GitHub, with your platform and version. App error reports carry a fixed message and up to ten app stack frames with relative file names. They leave out raw error messages, terminal content, prompts, command arguments, absolute paths and environment values. Each kind of error reports at most once every five minutes and 20 times an hour.

PostHog's code ships inside the app, so remote add-ons such as session replay can't load. The website (runpane.com) also records page views with PostHog.

To opt out, open **Settings → Privacy** and turn off **Allow product analytics**. Pane sends one identified opt-out event, then drops queued events and sends no more analytics, error reports included. An opt-out you made in an older version stays in effect.

## Agent permissions

Each agent has its own permission model: Claude Code uses `.claude/settings.json`, Codex has its own approval prompts, Aider has confirmation flags, and so on.

Pane's built-in agents start in their skip-the-prompts modes, so they can work without stopping to ask:

| Agent | Pane starts it as |
|-------|-------------------|
| Claude Code | `claude --dangerously-skip-permissions` |
| Codex | `codex --yolo` |
| Cursor | `cursor-agent --force --trust` |
| OpenCode | `opencode --auto` |
| Pi | `pi` (Pi runs tools without asking by default) |

The worktree keeps an agent's changes off your main branch, but the agent can still run any command your user can. Want an agent to ask first? Add it as a [custom command](/docs/custom-agents) with your own flags, like plain `claude`.

## Worktree isolation

Each agent starts in its own worktree folder. That sets where it works by default. It's a convention, and the operating system doesn't enforce it: an agent that uses another worktree's path can read it.

## Tailscale access security

The [Workspaces](/docs/runpane-workspaces) listener uses Tailscale Serve on port **8443**. Tailscale identifies each caller. Pick who gets in under **Settings → Remote Access → Tailscale → Who can connect**:

| Who can connect | Who gets in |
| --- | --- |
| **Only me** (default) | Devices signed in to your own Tailscale login. |
| **Everyone on tailnet** | People with untagged devices on the current tailnet. Pane warns that they could run agents and shell commands, and changes nothing until you click **Allow Everyone**. |
| **Off** | Nobody over this listener. Same as `runpane workspace disable`. QR-paired phones and connection codes still work. |

Tagged devices and devices shared in from another tailnet are always refused. So are browsers, including [runpane.com/app](https://runpane.com/app/). A browser connects with a connection code instead.

Passwords are optional and off by default. With one set, every Tailscale client needs it, your own computers included. Pane stores only a hash. Only someone at the host itself can change **Who can connect** or the password. `runpane workspace` sends the password the desktop app saved for that computer.

## Connection code security

> **Warning:** Treat a connection code like a password. Never paste one in a public channel.

The code-based daemon listens only on <Term id="loopback">loopback</Term>, `127.0.0.1:42137` by default. It never binds a public or LAN address. To reach it from another device, setup puts it behind <Term id="tailscale">Tailscale</Term> Serve on port **443**, you forward the port over <Term id="ssh">SSH</Term>, or you put your own HTTPS reverse proxy in front of it.

Every connection must carry the <Term id="bearer-token">bearer token</Term> inside the `pane-remote://...` code. The host makes the token during setup, and it gives full session control to whoever holds it.

**Encryption**: Tailscale wraps the connection in WireGuard, and Serve answers only on your tailnet, so nothing opens to the internet. The SSH tunnel uses SSH's own encryption.

**New codes and revoking**: Open **Settings → Remote Access → Tailscale → Set Up Host** and click **Create Connection Code**, or run `pane --remote-setup` on a headless host. The same screen lists each code-based client by device label. Disconnect or revoke any of them there, and a revoked token stops working immediately. Clients on the 8443 listener don't appear in this list.

**Data isolation**: Each data mode (**Current Pane Data** or **Isolated Daemon Data**) has its own database and worktree folders. Panes and Sessions stay in the mode they were made in.

For voice dictation in a remote browser, the host uses the voice keys in **Settings → Integrations → Voice transcription** or the host's environment variables. Live streaming uses Deepgram plus OpenRouter. Recorded transcription uses Fal plus OpenRouter.

## Reporting a vulnerability

Don't open a public issue, pull request or discussion. Email the security contacts, both if you can: **khazapar7@gmail.com** and **tyler@tbrown.io**. Include your Pane version and OS, what an attacker controls and gains, and the smallest steps that reproduce it. We'll reply by email and keep you updated until it's fixed.
