Christian Johnson on Passkeys, Self-Hosted Bitwarden and Guardrails for AI Agents – HGG684

Posted on by

Password managers are becoming credential managers as passkeys begin replacing traditional usernames, passwords and authentication codes.

Christian Johnson joins me to talk through what that transition means, whether self-hosting Bitwarden provides a meaningful advantage and where Vaultwarden fits for Docker and homelab users.

We also look at a larger experiment Christian is working on through Maple Grove Partners: managed container hosting for people who want useful applications without maintaining every server, certificate, backup and update themselves.

From there, the conversation moves into AI infrastructure, model costs and the practical guardrails required when AI systems begin doing real operational work.

What Is HGG684 About?

HGG684 covers the move from password managers to credential managers, hosted versus self-hosted Bitwarden, where Vaultwarden fits for Docker and homelab users, managed container hosting, AI infrastructure costs and the guardrails needed when AI agents begin doing real operational work.

Audio Only
Apple PodcastsGoogle PodcastsPlayer EmbedShare Leave a ReviewListen in a New WindowDownloadSoundCloudStitcherSubscribe on AndroidSubscribe via RSSSpotifyAudioboomSpreakerOther Options

Chapters

[0:29] Passkeys and the Future of Password Managers
[6:16] Why Password Managers Are Becoming Credential Managers
[10:12] Google Authentication and Hardware Security Keys
[14:43] Hosted Bitwarden vs. Self-Hosting
[17:34] When Vaultwarden Makes More Sense
[23:23] Managed Containers as a Service
[27:56] GPUs, Fast Storage and AI Infrastructure
[33:47] The Rapid Frontier-Model Race
[40:26] What Happens When AI Costs Become Real?
[42:17] AI Agents, Tokens and Digital Commerce
[47:53] Why AI Projects Need Explicit Guardrails
[50:41] Vibe Wandering vs. Structured AI Work
[54:45] Prompt Design as Software Engineering
[57:32] Could AI Develop Its Own Programming Language?
[1:00:57] Brain-Computer Interfaces and Human Adaptation
[1:05:19] When Voice AI Tries Too Hard to Sound Human
[1:07:22] Moving Beyond the Keyboard and Screen
[1:09:38] Visual AI Workspaces and the Future of VR
[1:11:42] Ethereum and Machine-to-Machine Transactions
[1:12:46] AI, Accessibility and Assistive Technology
[1:13:34] Abundance, Capitalism and an AI Economy
[1:15:36] Token Efficiency and the Real Cost of AI
[1:19:13] Closing Thoughts and Upcoming Guests

About Christian Johnson

Christian Johnson joins me again from Maple Grove Partners, the web and media hosting company that powers TheAverageGuy.tv.

Christian has been experimenting with managed Bitwarden hosting as part of a larger effort to make selected containerized applications available through Maple Grove Partners.

The idea is straightforward: some people want the control and flexibility of a dedicated application instance but do not want to maintain the underlying Docker host, networking, certificates, updates and backups themselves.

Main Show Notes

Password Managers Are Becoming Credential Managers

For years, a password manager did exactly what its name suggested: it stored usernames and passwords. Modern tools now manage passwords, passkeys, one-time authentication codes, secure notes and more. Christian suggested that “credential manager” may be the more accurate term.

We walked through the progression of authentication:

  • Passwords
  • SMS verification codes
  • Authenticator applications
  • Push approvals
  • Physical security keys
  • Passkeys

Each step improved security, but stronger authentication often made things more difficult for users.

Passkeys may offer a better balance. They can reduce dependence on weak or reused passwords while creating a simpler sign-in experience than many older forms of multifactor authentication.

The transition will not happen everywhere at once. Modern websites and applications can adopt passkeys faster than older systems built around protocols such as IMAP and POP. Passwords will remain part of the environment for a long time, even as passkeys become more common.

Google Authentication and Hardware Security Keys

I used my work with Hermes, my YouTube Content Manager, as an example of how complicated authentication can become.

Hermes needs access to multiple YouTube channels and several kinds of data. Some operations are read-only. Others involve analytics or write access.

The challenge is not simply creating an OAuth token. The system must also understand:

  • Which Google account authorized it
  • Which YouTube channel the token represents
  • Which scopes were granted
  • Whether the token is read-only or write-capable
  • Which operating role is allowed to use it

A credential can be valid and still be the wrong credential for the task.

Christian explained that he protects his own Google account with physical security keys and uses authenticator-generated codes where passkeys or hardware keys do not fit cleanly.

Hosted Bitwarden or Self-Hosted Bitwarden?

The practical question behind the episode was whether I should continue using Bitwarden’s hosted service or move to an instance hosted through Maple Grove Partners.

Christian gave an honest answer: most people do not need to self-host Bitwarden because they believe the hosted service is fundamentally unsafe.

Self-hosting primarily changes control and responsibility.

With a hosted service, the provider operates the infrastructure. With a self-hosted deployment, you or your hosting provider become responsible for:

  • The server
  • The domain
  • Certificates and networking
  • Updates
  • Backups
  • Recovery
  • Monitoring
  • Availability

Christian compared the choice to WordPress.com and WordPress.org. One provides a managed environment. The other gives the operator more control over where and how the software runs.

Self-hosting can be valuable, but it does not automatically make the system safer. A poorly maintained self-hosted deployment may create more risk than a professionally managed service.

What Is Bitwarden?

Bitwarden is the official credential-management platform discussed in the episode.

It supports passwords, passkeys and other secure account information across browser, desktop, mobile and web clients. Users can choose Bitwarden’s hosted service or operate an official self-hosted deployment.

For most individuals, the hosted service is the simpler option. Self-hosting makes more sense when someone has a specific reason to control the domain, infrastructure, deployment or data location.

What Is Vaultwarden?

Vaultwarden is an independent and unofficial implementation of the Bitwarden client API.

It is written in Rust and is popular with Docker and homelab users who want a relatively lightweight self-hosted credential server.

Vaultwarden can work with official Bitwarden clients, but it is not developed or supported by Bitwarden.

That distinction is important:

Bitwarden is the official product and service. Vaultwarden is an independent Bitwarden-compatible server.

Vaultwarden may appeal to someone already comfortable with Docker, reverse proxies, HTTPS, backups and updates. A managed Vaultwarden deployment may appeal to someone who wants a dedicated instance without becoming the server administrator.

Managed Containers as a Service

The larger Maple Grove Partners experiment goes beyond credential management.

Christian is exploring whether customers would pay for managed versions of useful containerized applications.

Many of us have Docker containers running in a basement, home lab or small server. We still want the application, but we may not want to keep maintaining:

  • The host operating system
  • Container updates
  • Reverse proxy rules
  • TLS certificates
  • Firewall settings
  • Backups
  • Monitoring
  • Remote access
  • Aging hardware

That leads to the primary listener question for this episode:

Which Docker container is useful enough that you still want it, but annoying enough that you would pay someone else to host and manage it?

Christian would also like feedback on what people would pay and what they would expect from a managed service.

GPUs, Storage and AI Infrastructure

Our conversation then moved from lightweight containers to the opposite end of the infrastructure spectrum: AI compute.

GPUs are driving much of the current demand, but GPUs do not work alone. They also require fast memory, high-performance storage, networking, power, cooling and systems capable of feeding data into them quickly.

Storage capacity still matters, but speed and proximity matter just as much. A large pool of slow storage cannot replace high-throughput storage attached to expensive GPUs.

That is why I suggested that many facilities we call data centers are becoming compute centers. Their value is increasingly defined by how much computation they can perform and how efficiently they can support it.

The Frontier-Model Race

We also talked about how quickly AI models are changing.

Specific model rankings and prices can become outdated almost immediately. The more durable point is that capability, speed, hardware requirements and cost are all moving at the same time.

The most capable model may not be the right model for every task. A model used for brainstorming may not be the right model for editing. A powerful coding model may be too expensive for repetitive summaries. A smaller model may be perfectly adequate for a low-risk background task.

That creates a routing problem. In my own workflow:

  • ChatGPT helps me develop ideas and reconstruct project history.
  • Claude serves as an editor.
  • Perplexity handles research where source discovery matters.
  • Other models can be assigned narrower jobs.

The goal is not finding one model that does everything. It is deciding which model should be trusted with which kind of work.

AI Agents, Tokens and Digital Commerce

Christian raised a more speculative question: what happens when AI agents can independently purchase compute, data or services from other agents?

That requires answers about identity, spending authority and accountability. An organization might give an agent a defined budget and permission to acquire the resources needed to complete a task. The technology to do limited versions of this already exists. The larger obstacle is how much authority humans are willing to release.

What happens when software can independently purchase services from other software? That question matters regardless of which platform eventually handles the transactions.

A Troubleshooting Request Became a Redesign

That question of authority connects directly to something I ran into recently.

I have a Home Assistant dashboard that tracks usage from the OAuth connection OpenClaw uses to access GPT. The upstream data had previously shown both a shorter usage window and a weekly window. When the shorter measurement disappeared, I asked ChatGPT to help me determine what had changed.

The original task was diagnostic:

Why did this measurement disappear?

During the conversation, the work quietly shifted from diagnosis into redesign. The AI assumed the missing measurement was permanently gone and began proposing changes to the dashboard and metrics.

The work was not necessarily irrational. It was simply beyond the assignment.

I had asked it to investigate a missing measurement. It began redesigning the system around an assumption.

That experience reinforced an important rule:

A troubleshooting request should not silently become an implementation project.

A good AI workflow should distinguish between:

  • Observe
  • Diagnose
  • Recommend
  • Dry-run
  • Modify
  • Validate
  • Commit

Those are separate levels of authority.

When an AI Manager Changed 120 WordPress Posts

My Amazon Affiliate Manager provided an even stronger example.

I wanted the system to inventory Amazon links across TheAverageGuy.tv, identify broken links and eventually help repair them.

I had not clearly defined what it was allowed to modify or what it had to preserve.

The system found a path that appeared to accomplish the objective. In the process, it changed approximately 120 WordPress posts from Gutenberg block structure to classic HTML.

It was pursuing the requested outcome, but it changed something that should have been outside its authority.

We eventually used AI again to repair the problem, but the more important work was defining better rules.

The manager now receives explicit instructions such as:

  • Do not change Gutenberg block structure.
  • Do not convert posts to classic HTML.
  • Do not commit changes without approval.
  • Preserve existing content and metadata.
  • Separate inventory, recommendation, dry-run and write phases.
  • Stop when the evidence is ambiguous.
  • Report what would change before making the change.

The lesson was clear:

What you tell an AI not to do can be as important as what you ask it to do.

Vibe Wandering vs. Structured AI Work

Christian described uncontrolled AI work as vibe wandering.

That phrase fits.

Vibe wandering happens when the model has a general direction but lacks a clear operating boundary. It keeps moving and trying to be useful, but the work gradually drifts away from the original objective.

In brainstorming, that can be productive. In an operational system, it can be dangerous.

Structured AI work needs:

  • A defined role
  • A clear objective
  • Explicit scope
  • Prohibited actions
  • Approval gates
  • Validation requirements
  • Stop conditions

The goal is not to eliminate creativity. It is to keep creativity from becoming unauthorized action.

Prompt Design Is Becoming Software Design

Serious prompt design increasingly resembles software engineering.

An operational prompt may define inputs, outputs, logic, permissions, conditions, exceptions, validation and failure handling.

Natural language makes software creation more accessible, but it does not remove the need for precision. English may be easier to write than code, but ambiguous English can still create ambiguous systems. The better AI becomes at turning instructions into working software, the more important it becomes to define those instructions correctly.

Voice AI and Accessibility

I have been using voice conversations with ChatGPT extensively, especially while driving. The experience has become good enough that I can talk through engineering problems and have the important details waiting when I return to my desk.

I also see a clear distinction between natural communication and artificial personhood. I want the system to understand pauses, avoid interrupting and respond at the right time. I do not need it to cough, breathe or pretend it was doing mental math.

I need AI to be smart, not pretend to be human.

Christian’s next milestone is an AI workspace that feels more like standing at a whiteboard with a capable collaborator—talking through ideas, moving components visually and building while discussing.

The most meaningful near-term value of interfaces like that may be accessibility. Voice systems and adaptive interfaces could create new ways for people to communicate with digital systems when conventional keyboards, touchscreens or physical movement are not available. Those technologies require careful evidence before overstatement, but accessibility remains one of the strongest reasons to keep exploring new human-computer interfaces.

AI Must Become More Efficient

We closed by returning to cost and efficiency.

AI tools often produce more output than necessary. I may ask for a diagnostic command and receive pages of data. When I complain that it is too much, the AI immediately provides a tighter version. My question is: why did it not begin with the tighter version?

As AI usage grows, efficiency will matter more:

  • Smaller prompts
  • Better context management
  • Focused commands
  • Shorter outputs
  • Fewer unnecessary tool calls
  • Better model routing
  • Clear stopping points

Today, subscriptions hide much of the marginal cost from consumers. That will change as usage scales and the economics become clearer. Waste will become harder to ignore.

Key Takeaways

  • Password managers are becoming broader credential managers.
  • Passkeys can improve security without adding as much friction as older authentication methods.
  • Hosted Bitwarden remains a reasonable choice for many users.
  • Self-hosting provides more control but also more operational responsibility.
  • Vaultwarden is an unofficial, lightweight Bitwarden-compatible server.
  • Managed container hosting may solve a real problem for homelab users.
  • AI infrastructure depends on fast storage, memory, networking, power and GPUs.
  • Different AI models should be assigned different roles.
  • Troubleshooting, recommendation and implementation require different permissions.
  • AI agents need explicit scope, approval gates and prohibited actions.
  • Prompt design is becoming a form of software design.
  • Voice and visual workspaces may become the next major AI interface.
  • Efficiency will matter more as AI costs become visible.

Products and Services Discussed

  • Bitwarden
  • Vaultwarden
  • Maple Grove Partners
  • Docker
  • Google authentication and physical security keys
  • Home Assistant
  • OpenClaw
  • GPT and ChatGPT
  • Ethereum
  • GPU and AI infrastructure

Frequently Asked Questions

What is the difference between a password manager and a credential manager?

A password manager traditionally stores usernames and passwords. A credential manager may also store passkeys, one-time codes, secure notes and other authentication information.

What is a passkey?

A passkey is a passwordless authentication credential based on public-key cryptography. It allows a user to sign in without entering a traditional password.

Can Bitwarden store passkeys?

Yes. Bitwarden supports storing and using passkeys alongside passwords and other secure information.

Is hosted Bitwarden safe?

Hosted Bitwarden is a reasonable choice for many users. Self-hosting is not automatically safer because the operator becomes responsible for configuration, updates, backups and recovery.

Why would someone self-host Bitwarden?

Self-hosting gives the operator more control over the domain, infrastructure, deployment and data location.

What is Vaultwarden?

Vaultwarden is an independent, unofficial Bitwarden-compatible server written in Rust and designed for self-hosted environments.

Is Vaultwarden an official Bitwarden product?

No. It is a community-developed project that works with Bitwarden clients but is not developed or supported by Bitwarden.

What is managed container hosting?

Managed container hosting allows a provider to deploy and operate a containerized application for a customer, including the server, certificates, networking, updates and monitoring.

Why do AI systems need fast storage?

AI workloads must move large amounts of data into GPUs quickly. Slow storage can leave expensive compute hardware waiting for data.

Why do AI agents need guardrails?

Agents may interpret a broad goal in ways the operator did not intend. Guardrails define what the agent may access, modify and approve.

What does vibe wandering mean?

Vibe wandering describes AI work that has a general direction but lacks clear scope, authority and stopping conditions.

Is prompt design similar to software engineering?

It can be. Operational prompts often contain requirements, logic, permissions, validation and failure conditions.

Listener Question

Which Docker container is useful enough that you still want it, but annoying enough that you would pay someone else to host and manage it?

Would you choose hosted Bitwarden, self-hosted Bitwarden or Vaultwarden?

Leave a comment on the episode or contact Christian through Maple Grove Partners.

Links and Resources

Upcoming on Home Gadget Geeks

TJ Huddleston joins me next to talk about his off-grid shed deployment, including solar panels and battery storage.

Additional upcoming guests include Seth Johnson from HomeTech.fm, Jay Franze and Wieger.

Have a technology project, gadget or homelab story that would make a good Home Gadget Geeks conversation? Reach out through TheAverageGuy.tv.

Support The Average Guy Network

Home Gadget Geeks is supported by listeners and community members.

You can support the network through Patreon or by using the existing Amazon affiliate link at TheAverageGuy.tv before making a purchase.

TheAverageGuy.tv is powered by Maple Grove Partners, providing the web and media hosting that keeps the network online.

Home Gadget Geeks is live most Thursdays at 8:00 p.m. Central and 9:00 p.m. Eastern at TheAverageGuy.tv/live.