Portal Expose Agent Skill

Expose an App with Portal

Portal publishes a service that is already running on the user's machine. It does not build the app or move it to a cloud host. Treat a successful tunnel as dependent on both the local app and the Portal process or agent remaining available.

Read references/portal-cli.md when choosing commands or persistent-agent configuration. Read references/x402.md whenever the user requests x402 or a paid route. Read references/safety-and-verification.md before exposing a nontrivial project, a service with authentication, or any non-HTTP port.

Choose the Mode

Use the smallest mode that satisfies the request:

• Temporary web preview: portal expose <target>.
• Trusted static directory or HTML entry: portal expose --serve <path>.
• Multiple local HTTP services under one URL: repeat --http-route.
• Paid HTTP path: routed HTTP with an explicit x402 payment contract; never enable payment implicitly.
• Durable tunnel that should survive terminal or login restarts: an explicit portal agent config and managed service.
• Session-owned durable tunnel without an OS service: portal agent run --foreground.
• Raw TCP or UDP: only when the user explicitly requests that transport and names the target.

Default to a temporary preview when the user says only "share", "preview", or "deploy locally". Do not install an OS service unless the user asks for a persistent, managed, or restart-surviving tunnel and accepts that portal agent run without --foreground installs a per-user launchd or systemd unit.

Workflow

1. Inspect the Project

• Read the applicable repository instructions before running or changing anything.
• Determine the app directory, start command, expected protocol, loopback target, and a meaningful health path.
• Prefer declared scripts and documented ports over guessing from process lists.
• Do not expose a port merely because it is listening. Tie it to the requested app.
• If the project is already running, preserve its process. If it is not running and deployment was requested, start it with the project's normal command and retain the terminal/session handle.

Ask one concise question only when the target, desired lifetime, or transport cannot be discovered safely. An explicit request to deploy, publish, expose, tunnel, or share authorizes creating the public tunnel for the named app; it does not authorize exposing adjacent services.

For x402, do not guess the protected path, payment methods, amount, network, recipient, or network-specific asset. Collect any missing consequential value before building the command or config. Treat an omitted method list as charging every method on the route and confirm that scope when it was not explicit.

2. Verify the Local Service

• Wait for the app's real readiness signal, not only for the process to exist.
• Make a bounded local request to the selected target. For HTTP, record the URL and status. For TCP/UDP, use a protocol-appropriate check that does not mutate application data.
• Stop before opening a tunnel if the local health check fails.
• Warn and require explicit direction before exposing databases, container daemons, debug consoles, unauthenticated admin panels, or services containing sensitive data.
• Before opening the tunnel, say that the public hostname is listed on participating relays and visible via portal list unless the user asked for --hide.

3. Check Portal

• Run portal version when portal is available.
• If Portal is missing, present the official install method and request approval before running it because installation writes outside the project. Never execute an installer from an unknown relay or third-party URL.
• Do not assume a hard-coded latest release or stale flags. Use the installed version and the checked-in Portal reference as the compatibility baseline.

4. Build the Command or Agent Config

• Use loopback targets such as 127.0.0.1:<port> unless the project explicitly needs another address.
• Use the user's requested name. Otherwise omit --name for a temporary preview or derive a stable DNS-label-safe name for a persistent tunnel.
• For portal expose, always pass an absolute --identity-path outside the repository. The CLI default is identity.json in the process working directory and that file contains private key material. For portal agent, omit identity_path so the agent stores identity under its state directory; if you set the field, use an absolute path outside the repository.
• Never print or commit identity JSON, control tokens, facilitator tokens, or wallet secrets.
• With a user-selected relay on portal expose, pass --relays <https-url> --discovery=false. In persistent mode those flags are not accepted on portal agent run; put relays = ["https://..."] and discovery = false on the [[tunnels]] entry instead.
• The MITM self-probe always runs. Without --ban-mitm / ban_mitm = true, a suspected TLS termination is only logged and the tunnel keeps serving. Do not claim the default path blocks a relay. Add --ban-mitm only when the user wants fail-closed handling. There is no flag that disables the probe.
• Never add TCP, UDP, multi-hop, payment, or public metadata flags that the user did not request. --hide is the exception for listing: mention the default public listing, then add --hide or hide = true only when the user wants the tunnel unlisted.
• For a paid route, follow references/x402.md. Keep payment policy on the smallest requested path, use an explicit network, and never place wallet or facilitator secrets in a command, log, committed file, or final response.

Before executing, show the exact public target and any important exposure consequence when it is not already obvious from the user's request.

5. Start and Observe the Tunnel

• Run a temporary portal expose in a foreground PTY or managed long-running command session. Do not hide it behind an untracked nohup process.
• For persistent mode, inspect any existing agent config and running service first. run, restart, and stop are service-wide: they affect every [[tunnels]] entry that the selected service owns. Reuse and merge the existing config when the same agent should keep other tunnels. An isolated second agent needs its own config, service_name, state_dir, and loopback control_addr. Changing only service_name still shares the default state directory and 127.0.0.1:4018. Do not stop or replace an agent that already owns unrelated tunnels.
• Create or update only the selected agent config, then start it with portal agent run --config <path> after the user accepts OS-service installation, or portal agent run --foreground --config <path> when the current session should own the process. --foreground opens the interactive dashboard when stdin and stdout are TTYs. Run that command in a non-TTY managed session so logs stay capturable and the TUI does not start.
• Do not run portal agent dashboard. It is an interactive TUI. Give the user that command in the handoff.
• Capture bounded output. Redact tokens, identity material, signed payloads, and credentials.
• HTTP tunnels are ready when a public URL is emitted. Raw TCP/UDP tunnels log raw transport endpoints allocated with tcp_addr and/or udp_addr instead of service ready at <URL>. Do not wait for an HTTPS URL on a raw transport.

6. Verify the Public Endpoint

• For HTTP, make a bounded HTTPS request to every public URL being handed off. A deliberately authenticated app may return 401 or 403; explain that as reachable but protected. Treat unexpected 5xx, TLS errors, or a Portal error page as a failed deployment.
• For each paid route, make an unpaid request with a protected method and require 402 Payment Required plus a payment-requirements header. Compare the returned network, asset, recipient, amount, and resource with the requested policy. Verify the method scope by requesting an intentionally unprotected method when one exists. Never spend funds merely to verify configuration.
• For raw TCP or UDP, protocol-probe the allocated tcp_addr/udp_addr without mutating application data. A successful local port open is not enough.
• When a browser-capable tool is available and the app has UI, load the primary page and check for an obvious render or runtime failure. Do not log in or submit data unless the user requested it.
• Re-check the local health endpoint if the public request fails so the handoff distinguishes app failure from tunnel or relay failure.

7. Hand Off the Result

Report:

• Deployment mode and exact local target.
• Public URL or allocated raw endpoint, and the verified status.
• Whether the tunnel is listed on public relays or hidden with --hide.
• Whether MITM handling is detect-only or --ban-mitm.
• For x402, the protected paths and methods, human amount, network, public recipient, facilitator mode, and whether the unpaid 402 challenge was verified. State explicitly when settlement was not tested.
• The identity path and that it must stay out of version control.
• The app and Portal process/session or OS-service ownership.
• The exact stop or restart command, and whether that command affects other tunnels on the same agent.
• Anything that remains temporary, unavailable, or unverified.

Do not call the result permanent when the local machine, app process, or foreground tunnel must remain running.

Failure Rules

• Local app unhealthy: stop before exposing it and report the failing check.
• Portal absent and installation not approved: provide the official command without executing it.
• No ready public URL or allocated raw endpoint: keep the bounded diagnostic output and report the relay/tunnel failure.
• Paid route returns anything other than the expected 402 challenge: do not describe it as protected or hand it off as ready. Stop only the tunnel created by this workflow, preserve bounded diagnostics, and report the policy mismatch.
• MITM self-probe warning without --ban-mitm: report the warning and offer --ban-mitm; do not claim the relay was blocked.
• Requested name unavailable: offer an auto-generated or alternative name; do not silently hijack another identity.
• Existing agent owns other tunnels: do not stop or replace it to publish this app.
• Cancellation: stop only processes started by this workflow, unless the user explicitly asks to stop an existing app or agent.