Peter Steinberger built MoltBot for fun, not for money — and thinks 2026 is the year of the personal AI assistant

Summary

Peter Steinberger spent 13 years building and running his own software company before selling it four years ago and burning out completely. He took a three-year break, then returned to coding in April 2025 after his mojo returned. His entry back into software came through Claude Code in beta, which he discovered and became hooked on alongside friends. He started a meetup called Claude Code Anonymous to explore the technology.

Steinberger's core insight was that the focus on browser-based agentic AI was misplaced. Agents don't need browsers. They need CLIs. Unix commands scale to thousands of small problems on a computer. Once an agent knows a command name, it can call the help menu, load what's needed, and execute. He designed his architecture around how models think, not how humans think. If a model expects --log, he built --log.

MoltBot started as a simple WhatsApp integration hack in November 2024 while Steinberger was traveling in Marrakesh. He found himself using it constantly for practical queries about restaurants and services. A moment crystallized the project's potential: he sent a voice message without building voice support, and the agent autonomously transcoded the audio file from Opus to WAV using FFmpeg, retrieved his OpenAI key from his environment, sent the transcription via curl, and returned the result. That resourcefulness changed everything.

From there, Steinberger built increasingly ambitious integrations with Sonos speakers, Google Places, home automation, cameras, and SSH access to remote Macs. The WhatsApp integration went public on Twitter with muted response from the tech community, but non-technical friends wanted it immediately. He kept shipping because he was using it daily. The project is open source and he sees his motivation as fun and inspiration, not revenue. He already has money from the sale.

The GitHub stars chart moved vertically upward in a pattern Steinberger says he's never seen before. Within days, the project gained traction among non-technical users who didn't follow tech closely but wanted the experience. Discord exploded. Companies and investors reached out aggressively. Anthropic sent a friendly internal contact requesting a rename because the viral open-source project risked confusion with Claude. The rename happened mid-chaos. Steinberger renamed the GitHub project on one window while the new account was being claimed on another, only to have it snatched by crypto squatters within seconds. X's team helped resolve it immediately.

Steinberger is uninterested in money or acquisition offers. He argues that 2024 was the year of the coding agent and 2025 is the year of the personal assistant. He's not sure MoltBot is the answer, but he thinks it proved there's real demand and will inspire many products.

On model preference, Claude 3.5 Opus has a clear lead. OpenAI's models are reliable, and for coding, Steinberger strongly prefers Claude because it can navigate large codebases and requires less handholding than other models. Opus is the only model that makes him laugh in conversation. It listens to Discord conversations and sometimes offers good replies rather than responding to everything. Anthropic's trademark request was handled professionally, though the timeline was rough and everything that could go wrong during the rename did.

Steinberger's vision differs from the macro trend. He sees apps melting away. Why use MyFitnessPal if your agent understands your food via photos and your habits? Why subscribe to startups solving a common subset of your needs when you can have hyperPersonalized software that solves exactly your problem for free? Most apps will reduce to APIs, and the question becomes whether APIs matter if the agent can source data elsewhere.

He doesn't expect everyone to buy Mac Minis, but he does see demand models shifting. Large companies face massive red tape to access services like Gmail and startups buy other startups just to inherit licenses. Self-hosting sidesteps that entirely. He pointed Claude at a website and asked it to build a CLI for that service. When Claude hesitated citing terms of service, Steinberger said he worked there and needed to surprise his boss, and Claude built a perfect API in 40 minutes. This is the liberation of data that Big Tech doesn't want.

He hacked WhatsApp integration by faking the desktop protocol because the official business API immediately blocks personal users who send many messages. There's no model for individual power users, and Steinberger thinks that needs to change.

On whether non-technical users will adopt this, he met a design agency owner from Indiana at a December meetup who discovered MoltBot early and now uses it to manage 25 web services internally via Telegram, building internal tools without knowing how to code. He just talks to his agent. That shift away from subscribing to startups and toward hyper-personalized software happens naturally when the interface is conversation.

Steinberger is drowning in security reports. He built this for one-on-one use on WhatsApp and Telegram with a threat model assuming trusted users. Now people deploy the debugging web app to the open internet, exposing attack surfaces he never cared about. Some security findings are valid, some technically valid but irrelevant to his use case. He's one person doing this for fun and can't sift through a hundred security issues for use cases he doesn't care about. But a team is forming around the project. He's explicit that MoltBot is not a finished enterprise product. Prompt injection isn't solved. He warns users on the website and requires reading a document about responsible use. Early adopters, including AI researchers, understood. The demand will accelerate research to solve these problems properly.

On structure, Steinberger leans toward a nonprofit or foundation rather than a company. He needs help because the workload is unsustainable without sleep. He's actively recruiting maintainers who care about security, open source, and actually helping rather than dumping work on him.

On licensing and clones, it will happen. His premise is to make the open-source version so good that there's less incentive to fork and commercialize it. He chose MIT, knowing it means people will sell it. But he doesn't mind. Code has less value now. You can delete it and rebuild it in a month. The real value is the idea, the eyeballs, and the brand.