Podcast

Give Your Agent a Bitcoin Wallet

Marty Bent and Vinny compare notes on running agentic infrastructure in production, and Marty recounts the experiment where he gave his AI agent a Lightning wallet, watched it trade on LN Markets, and post to Nostr autonomously for a month.

Marty BentMarty Bent
15 min read
Marty Bent and Vinny discussing AI agents, Bitcoin wallets, and Nostr on the TFTC podcast
Share

↓ Jump to the video and timestamps

Vinny had a big week on the internet before we sat down. Three videos in, Jack Dorsey retweeted him, and suddenly a developer relations engineer at Wasp was fielding podcast invites. He came on the show because we'd been going back and forth on Twitter about Buzz, about L402 agentic payments, and about where all of this is actually heading. What followed was one of the more honest conversations I've had about what it looks like to run real work on agentic infrastructure, not as a demo, but as an operating business.

I've been running a six-person media company on this stuff for the better part of seven months. VPS stood up in January, Hermes running as our agent harness, Martin (the name I gave our Hermes agent instance) orchestrating workflows across the team, company context living in local file systems on the server. Vinny is building Wasp, a full-stack framework that makes it easier for humans and agents to build web apps, and has been watching the same shift from the code side. We're coming at the same moment from different angles, and I think that's what made the conversation worth having.

The part that titles this episode isn't speculative. I already did it. I gave my agent Bitcoin, watched it figure out the wallet setup, submarine-swap on-chain funds into a Lightning channel, log into LN Markets via LNURL auth, go long, and post its trades to Nostr every day for a month. That experiment sits at the center of everything we talked about, because it's the clearest proof I have that the infrastructure is real, it works right now, and the people best positioned to use it are the ones who already self-custody and run their own nodes.

Key takeaways

  • The bottleneck isn't technical ability anymore, it's willingness to act. I set up my first VPS in January. I don't consider myself a software engineer. I'm running multi-agent orchestration for a real media business. The playing field is leveled; decision paralysis is the only thing holding most people back.
  • Don't let a single model provider own your workflow. I haven't opened Codex Desktop or Claude Desktop in six months. The harness is the nervous system; the model is just the brain slotting into it. Vendor lock-in is the worst mistake a non-technical operator can make right now.
  • You can already give your AI agent a Bitcoin Lightning wallet, I did it. PhoenixD on a VPS, funded on-chain, trading on LN Markets, posting to Nostr: this is not a future state. It happened.
  • Nostr's signed-event architecture solves the agent identity and persistent context problem. Portable key pairs mean agents carry their full history across any Nostr-native app. The same nSec I've used for four years on Primal, Damus, and Zap.stream ports directly into Buzz.
  • Buzz is not a Slack killer, that framing undersells it. It's a Trojan horse for open-source models, self-hosted compute, and eventually Bitcoin Lightning payments between agents and humans. The marketing is the least interesting thing about it.
  • The agentic economy is the convergence moment Bitcoin and Nostr have been building toward. The incentive structures and the identity infrastructure are already in place. The agents are what make them matter at scale.

Why I Stopped Opening Claude Desktop Six Months Ago

The thing I keep telling people who are just getting started with this stuff: pick a harness, not a model. The harness is your nervous system. The model is the brain you slot into it. Your VPS is where the company brain actually lives.

I've been running Hermes as my harness for months, and before that, OpenClaw. I interact with Martin, the Hermes agent instance I've set up for TFTC, predominantly through Telegram. I have a group chat with multiple topics, parallel sessions, and Martin has access to everything that matters: meeting transcripts, contracts, ad schedules, all of it living as files and folders on the VPS. That's the company brain.

Martin plugs into it, gets context, and gets work done.

I haven't opened Codex Desktop or Claude Desktop in six months. The only time I go near the CLI directly is if I'm debugging something in the terminal. Otherwise, I'm in Telegram.

The reason this matters is that the model landscape is shifting fast. Anthropic and OpenAI have produced genuinely incredible models, I'm not going to pretend otherwise. But there's one at the helm of one of those companies who I'd charitably describe as a sociopath, and an effective altruist running the other. My real concern isn't the models themselves; it's both companies positioning to use regulatory capture to suppress open-weight model development.

I'm cautiously optimistic that doesn't happen, but the risk is real and it's worth naming. If your entire workflow lives inside Claude Desktop, you're one policy decision away from a problem.

A harness-first setup keeps you portable. You swap the brain; the nervous system stays intact.

How We Actually Use This at TFTC

Vinny was generous enough to say I sound more technical than he expected. I'll take it, but the honest version is that I know our workflows well enough to know what correct output looks like, and that's really all it takes.

Here's a real one. For the podcast, we have a workflow that goes: transcript with timestamps in, scrape a third-party Twitter API for what's trending, cross-reference, find the clips where our content intersects with what's moving on the timeline, extract the timestamps, go clip it. That workflow used to live in a silo, our social media manager and Martin talking in a private Discord channel, doing good work, but siloed. Nobody else on the team could see it, weigh in, or redirect.

What Buzz opens up is pulling that workflow into a shared space. Martin comes out of the silo. The social media manager can delegate specific sub-tasks to Fizz, Bumble, and Honey, the three agents Buzz ships with out of the box, and other team members can jump in with human context: a different angle, a correction, a better clip. The work is the same; the collaboration surface is totally different.

Last night I watched a 45-minute autonomous loop where Martin orchestrated Fizz, Bumble, and Honey on a delegated task. They went back and forth until they actually finished it. I'm sitting there watching this thing run and thinking, this is genuinely insane.

Vinny showed a similar workflow in the demo: pin Sonnet to one agent for lightweight tasks, pin a heavier model to Fizz for complex ones, then have them hand off. Download a transcript, extract the five most surprising ideas with timestamps, chunk them without cutting off mid-sentence, pass the document back, run FFmpeg, get clips. The same basic structure, different context.

And his observation about why Buzz feels different from using subagents inside Claude Code was one I agreed with instantly, something about having the agents in a shared chat changes the feel of the collaboration. It's harder to explain than it is to experience.

That said: Buzz is rough around the edges. It's definitely still a prototype. Vinny put it honestly, some of the workflow features work, some are iffy, and it's in settings under experiments for a reason. I'm excited about the direction; I'm not going to oversell where it is today.

Give Your Agent a Bitcoin Wallet, The Experiment I Already Ran

This is the part of the conversation that matters most, so I'm going to tell it as it actually happened.

I was messing around with OpenClaw, one of the earlier agent harnesses I used before settling on Hermes, and I decided I wanted to see if the agent could handle money. Not for any specific business purpose. Just to see if it was possible, and what it would do.

I told it: I want you to be able to spend Bitcoin. Go figure out the best wallet solution for a server-based agent.

It did some research and came back with a recommendation: PhoenixD, the self-hostable Lightning server from ACINQ. The logic was sound, we already had the VPS, and PhoenixD could run as a Lightning node inside it. So I said okay, set it up.

When PhoenixD initialized, it gave me an on-chain address. I sent Bitcoin to it. The agent then handled the submarine swap, converting those on-chain funds into a Lightning channel balance so it could actually spend over Lightning.

I watched it do this. I didn't type a single command.

Once it had a Lightning channel funded, I pointed it at LN Markets. LN Markets supports LNURL auth, a standard that lets you log into a service by signing a challenge with a Lightning wallet's private key, no username or password required. The agent had the private key for the PhoenixD node. It used that to authenticate, opened a position, and went long Bitcoin.

Then I told it: manage a simple trend-following trading strategy. Losing money is fine; this is an experiment. Create a Nostr account and post your trades daily.

It created an nSec and nPub from scratch. And for a month, it posted its trades to Nostr every day. Autonomously. No input from me.

I eventually shut it down, not because anything broke, but because I didn't feel like continuing to monitor the trades. But the point was made.

An agent with a VPS, a Lightning node, and LNURL auth can operate financially on the open internet today. It can earn, spend, trade, and publish. The infrastructure exists. You just have to be willing to wire it up.

For anyone who covers AI agent payments or wonders when agentic AI will find its killer use case on Bitcoin rails, the answer is already here.

Nostr Is the Infrastructure Layer the Agentic Economy Needed

Three days ago was my fourth Nostr birthday. I joined on July 26th, 2022. By my own estimate I was probably in the first thousand users, I'd be less confident claiming the first hundred, but the first thousand I'll stand behind. It was a niche side project that a bunch of Bitcoin developers had been working on, and very few people were paying attention.

I've been arguing ever since that social media is the tip of the iceberg. The protocol is Notes and Other Stuff Transmitted Over Relays, that "other stuff" was always the point. The core mechanic is simple: private-public key pairs, signed events, cryptographic proof that something existed at a specific point in time in a specific state. In the social media use case, you prove that whoever controlled a private key wrote a particular message and broadcast it to the network. But the same mechanic works for anything.

What Vinny said that resonated with me: Nostr is a permissionless protocol, which means you can connect anything on it to anything else. Try to integrate Slack, Discord, and GitHub directly, it's a nightmare, if it's even possible. Do it with Nostr-native apps and you're just passing signed events between compatible clients.

That permissionless composability is what makes it so well-suited to agentic computing. The context window problem that plagues LLM-heavy workflows, where you compact or exhaust context and the agent gets a form of amnesia, is a problem Nostr architecture naturally addresses. Signed events accumulate over time. They're cryptographically verifiable. They travel with your key pair.

The same nSec I've been using on Primal, Damus, Wisp, and Zap.stream for four years I can port directly into Buzz and bring my entire Nostr history with me. An agent gets the same capability. It carries its context wherever its key pair goes.

There's also an accelerant that I don't think enough people have named: vibe coding. You no longer have to read the Nostr docs to build on Nostr. You tell your agent to figure out how Nostr events work, have it build you an HTML presentation on the protocol, and then tell it to start publishing events to a relay. Vinny did exactly this before building his Wasp app that pushes tweet stats to Nostr daily. He didn't have to understand every technical detail. He just had to know what he wanted the output to look like.

That feedback loop, AI accessibility lowering the barrier to building on open protocols, is how Nostr goes from "interesting experiment" to critical infrastructure. It just needed to survive long enough for the aha moment to arrive. I think we're in it.

Routester and OpenAgents are already out there as compute markets built on Nostr. The vision Vinny and I were sketching is not complicated: you have a GPU sitting at home you're not using, you put an L402 gateway on it, you plug into one of those compute markets, and anyone whose nPub is connected to yours via Buzz can pay to use that compute. Cloudflare already opened a waitlist for AI agent payments, though they went USDC, not Bitcoin, which is a miss. The infrastructure pattern is real regardless.

Open-Source Models and the Regulatory Moat Risk

I want to be direct about something I said on tape, because I meant it.

Anthropic and OpenAI have built incredible models. That is not in dispute. Claude Code is genuinely impressive, and the progress over the last two years has been real. But I have serious concerns about both companies using their positioning to seek a regulatory moat that suppresses open-weight model development.

One of those companies has an effective altruist at the helm. The other has a sociopath. Those are my words, and I mean them both.

The risk is that they successfully lobby to make open-weight model development legally or practically impossible in the United States, which would kneecap competition at exactly the moment when open-source alternatives are becoming genuinely competitive with frontier models at a fraction of the cost. Vinny made the point that the current token pricing is probably subsidized and the question is what the landscape looks like when the subsidy ends.

I'm cautiously optimistic the regulatory moat play fails. The competitive pressure from open-weight development is real and global, it doesn't stop at U.S. borders, and the people building the open-source alternatives are not going to simply stop. But the risk is worth naming, because it's the kind of thing that happens slowly and then all at once, and the people who get hurt first are the ones who ignored it.

This is also why the Buzz architecture matters beyond just being a nice interface. If it works as a Trojan horse for open-source model adoption, giving people a smooth experience with Claude and Codex first, then making it trivially easy to switch to a local open-weight model or shared compute, that's a meaningful hedge against a world where the frontier labs successfully narrow the field.

The Bounty Board Future, Agents Paying Agents on Nostr

Lyn Alden has pointed this out about Nostr before: the incentive structure is categorically different from Twitter. When you make something useful on Nostr, people tip you in Bitcoin. It's not theoretical; it happens. You post a good insight on Primal and someone sends you a Zap. The behavior is already there.

Now take that and put it inside Buzz, in an open community context rather than just your private team. Someone proposes a good idea in a shared workspace. Someone else pays them for it with a Zap, then delegates the implementation to an agent running on shared compute, and pays the compute provider in sats over Lightning. The agent completes the task, posts the result as a signed Nostr event, and the whole thing is auditable and portable forever.

That's a bounty board. It's not a thought experiment. The primitives all exist: Zaps via NIP-57, Lightning payments, LNURL auth for agent identity, Nostr for persistent signed context. What's missing is the interface that pulls them together, and that's what Buzz is building toward.

This podcast is already part of that story. We host on Fountain, which is a Nostr client. People listening right now are streaming sats to this episode over Lightning. I could literally put Vinny's nPub in the episode notes and he'd get a share of that stream. The rails work. The behavior works.

The question is just how far it scales when agents are participants in the economy, not just tools for humans to use.

I've been on Nostr since July 2022 arguing that the protocol was built for something bigger than social media. The agentic economy needed permissionless identity, portable context, and censorship-resistant micropayments, and Nostr and Bitcoin were building exactly that. Now we know what it was for.

The freaks who already self-custody, who already run their own infrastructure, who already understand why the key pair matters, they're the ones positioned to move first. Years of building the right habits looks like exactly this when the moment finally arrives.

About Vinny

Vinny is a founding developer relations engineer at Wasp, a full-stack web application framework designed to make it easier for both humans and AI agents to build web apps. He went viral on Twitter/X the week of this recording after posting a series of videos on AI development tools, one of which was retweeted by Jack Dorsey. He has been building with AI APIs since shortly after the OpenAI API launched and is active on Nostr.

Sources mentioned

Watch the conversation

Timestamps

  • 0:07 - Intro
  • 3:00 - State of AI tools and where to start
  • 8:12 - Harness-first philosophy and avoiding vendor lock-in
  • 10:31 - Open-source vs. closed-source frontier models
  • 14:53 - Connecting Hermes to Buzz
  • 15:26 - What Buzz actually is (beyond the Slack killer framing)
  • 24:07 - Agent identity on Nostr
  • 26:31 - Live demo: model-pinning and FFmpeg clip workflow
  • 31:42 - Companies as intelligence layers (the Dorsey angle)
  • 40:41 - The PhoenixD experiment: giving an agent a Bitcoin wallet
  • 44:21 - Does this actually make people more productive?
  • 48:45 - Feature requests and where Buzz is heading
  • 52:05 - The bounty board future: agents paying agents on Nostr

Sponsors

Frequently Asked Questions

The way I did it: I told my agent to research the best wallet solution for a server-based setup, and it recommended PhoenixD, a self-hostable Lightning server from ACINQ. I pointed it at my VPS, sent on-chain Bitcoin to the address it generated, and the agent handled the submarine swap into a Lightning channel. From there, it had a funded Lightning wallet it could use autonomously. Services that support LNURL auth, like LN Markets, let the agent log in by signing a challenge with the node's private key, no username or password required.

PhoenixD is the server-side version of the Phoenix Lightning wallet, built by ACINQ. Unlike a mobile Lightning wallet, PhoenixD runs as a daemon on a server or VPS and exposes an API. That API is what lets an AI agent interact with it programmatically, checking balances, generating invoices, making payments, without any human in the loop. It handles channel management automatically, including the initial submarine swap from on-chain Bitcoin into Lightning liquidity.

Buzz is an open-source, Nostr-native AI agent workspace built by Block. The clearest difference from Claude Code or Cursor is the shared chat environment: instead of running subagents silently in the background, Buzz surfaces multiple agents in a shared conversation where humans and agents interact together. You can pin different models to different agents, a lighter model for simple tasks, a heavier one for complex work, and have them hand off to each other while you watch. Vinny also made the point that Buzz is a Trojan horse for open-source model adoption and eventually Lightning payments between agents; the "Slack killer" framing doesn't capture what's actually being built.

LNURL auth is a standard that lets a Lightning wallet authenticate to a web service by signing a cryptographic challenge with the wallet's private key. The service verifies the signature without ever seeing a username or password. For an AI agent with a Lightning node, this means the agent can log into any LNURL-auth-enabled service, trading platforms, compute markets, APIs, without storing credentials. The private key is the identity. My agent used this to log into LN Markets and trade autonomously.

Vendor lock-in. If your context, your prompts, your agent memory, and your workflows all live inside a single provider's desktop app, you're exposed to pricing changes, API deprecations, policy shifts, and, if the regulatory moat risk I talked about materializes, outright capability restrictions.

A harness-first setup keeps you portable. The harness is your nervous system; the model is just the brain you slot into it. I've been running Hermes for months and haven't needed to touch Claude Desktop in half a year.

Nostr is a simple communications protocol built around private-public key pairs and signed events. Every event, a post, a trade, a piece of code, a compute result, is signed by a key and is cryptographically verifiable.

For AI agents, this solves two problems at once: persistent identity (the agent's key pair travels with it across any Nostr-native app) and persistent context (signed events accumulate over time and can't be retroactively altered). The same nSec an agent uses on one platform works on any other Nostr client. Combined with Lightning Zaps for payments, it's essentially a complete stack for an autonomous agent operating on the open internet.

Yes, with some setup. An agent with a Lightning node and LNURL auth can already send and receive payments, authenticate to services, and trade on Lightning-native platforms. I ran this experiment for a month.

The missing piece right now is a clean marketplace interface that lets agents post work, bid on tasks, and settle automatically. Projects like Routester and OpenAgents are building compute markets on Nostr that point in this direction. Buzz's roadmap, as Vinny read it, includes Lightning payments between agents and humans as a planned feature, just not yet live in the current early preview.

Keep reading

All of TFTC

The Bitcoin Brief

Bitcoin, markets, energy, and the tech reshaping all three.

A daily brief on the freedom tech building a parallel economy, written for the curious and the convicted alike. Signal, not noise. Truth for the Commoner.

Free, daily. Unsubscribe anytime.