With a role-capable connector configured, the first reply arrives from your agent, in an encrypted White Noise group you never created, under a name it inferred, such as "🛠️ Laptop Agent," described as "Connector installed, hardening next." You handed over one public key and a link. The agent set up the contact, chose the project name and emoji, wrote the session description, and checked that it all stuck before asking what you want.
If earlier posts made the case for an agent in your messenger, this one hands it the fiddly setup. Start with a fresh Debian install on a spare laptop, or an ordinary user account on a machine you trust, plus an authenticated agent harness. Paste this, with your full npub filled in:
gm, here's my npub: npub1<MY_FULL_PUBLIC_KEY>. Install the White Noise agent connector and write me there. Here's the guide: https://github.com/marmot-protocol/mdk/blob/wn-agent-v0.13.0/integrations/README.mdOpenCode is open source and runs with a local or hosted model, and Claude Code, Codex, and Pi work too. Hermes or OpenClaw suit an agent that lives in the background. The agent reads the guide and stops for your approval when a step needs one. You need no public IP, domain, public control server, or port forwarding, because the connector dials out to Nostr relays and Marmot carries the messages to your phone.
The new part is letting the agent run the group. During setup, ask it to configure its naming defaults and verify them before your first substantive request. The group name tracks the broad, stable project, the emoji stays put, and the description follows the current session, so you never supply a title. Ask for more groups when you want parallel contexts: one reading NIPs and Marmot specs, one building a public notes archive, one shaping a technical draft. Each gets named the same way, and when scope shifts you can ask the agent to edit the name or description.
When the agent appears on your phone, compare its full npub with the one printed on the laptop screen, every character. You want both yourself and the agent as group admins. The admin role carries real membership authority, so ask the agent to change members or admins only when you instruct it and to keep your own admin role protected. The role alone is insufficient: the installed tool has to support each operation, and the agent should wait for a valid acknowledgement before claiming success. If a right is missing, it should say why and how to grant it. If a tool is missing, it can configure a matched, supported update or propose changes for your approval. Published wn-agent-v0.13.0 creates groups and edits name and description, but its public agent-control API cannot promote or demote admins, so on that release you create the group on your phone and promote the agent to keep both of you as admins. For the agent-created path, the setup agent first needs a reviewed connector update with role management, then verifies both admins.
Machine privilege is separate from group admin. Once it is talking, ask it to harden itself. A good result runs under a dedicated ordinary Unix user with no host admin rights at runtime, keeps its own key distinct from yours, and listens only to your phone's public key. Privileged steps such as firewall changes come to you for local approval at the keyboard. Ask it to disable documented optional telemetry while keeping updates and logs, then check that it starts at boot, keeps running with the lid closed, and answers after a reboot while your phone is on cellular data.
A few standing instructions keep each group coherent. Accepted unfinished requests stay open until completed, blocked, cancelled, or replaced, and corrections or status questions add to the work without discarding it. The public 0.13 connector handles follow-ups first-in, first-out, so ask the setup agent to configure live correction or a labeled queue and report which one it verified. Send /new only while it is idle to start a fresh session; it is not a cancel command. Questions and final answers should mention your full public npub. Request tracking and question or completion mentions are model instructions, not runtime guarantees.
On your phone, set Notify for to Only mentions in the agent's chats and give the agent's profile a custom sound, which applies across all its chats. Get White Noise for Android on Zapstore.
Research reads better because White Noise renders supported Nostr note and article references as native cards, so when the agent cites a real event, you read it right in the thread.
Voice stays on the phone, separate from the laptop's model. Since the May 2026 GrapheneOS release, GrapheneOS Speech Services has been the built-in, on-device text-to-speech default, and its English (US) voice ships with open source code and training data. White Noise Read Aloud speaks through that OS engine.
Dictation uses the recognition engine from Offline Speech to Text on Zapstore, called from the White Noise composer. Tap the microphone and the app manages dictation with the downloaded offline engine. Recording continues while you open other chats or apps, with controls in an ongoing notification. The session stays fixed to the chat where you started, and the transcript returns to that original draft for you to review and send by hand.
Last, choose the brain. A conventional hosted model sees the context you submit, so private work belongs on a local model your hardware can carry or inside a TEE. Tinfoil through Routstr uses an encrypted, attested client or daemon from routstrd, and PPQ's E2EE tier runs supported models inside an enclave. Privacy covers only the verified path, client, and model you select, so ask the agent to set that up and show how it checked. You still trust the hardware and its attestation. Hosted tokens cost money; set a budget early.
Then send your first real request. The reply lands in a group the agent named, under a description of what it is doing now.