TL;DR. Nobody has won the personal AI agent race. The five most-watched products each fail at something different. The deeper reason is that almost everyone is solving the wrong problem. The bottleneck is not the model, and it is not the hardware. It is whether the agent is wired into your workflow with access to your actual data. That is the problem Ostler is built around.
Nobody has won yet
Peter Yang’s Creator Economy newsletter recently tested five of the most-watched personal AI agents against ten capabilities a real personal agent needs: managing email, calendar and documents; running recurring tasks; remembering you; working on web and mobile; voice; personality; computer control; reliability; security (source).
The conclusion is blunt. Nobody checks every box.
- OpenClaw is the most flexible and the most powerful, with strong messaging and voice. The cost is reliability. Roughly a tenth of the author’s time with it is spent keeping it working.
- Hermes trades flexibility for stability. It is more dependable, communicates tasks more clearly and runs workflows on its own. It feels less alive than its rivals, but it stays up.
- Claude Code has the best personality and the strongest reasoning, and it is a pleasure to work with. It also has rate limits and uptime issues that get in the way of relying on it.
- Codex ships a beautiful desktop app, generous usage and the best browser and computer control of the five. It has no mobile app, which the author estimates removes about eighty per cent of how he would actually want to use an agent.
- Gemini is best positioned for the mainstream because it already lives inside Google Workspace and handles voice and video well. The hole is that it cannot edit Google Docs, Sheets or Slides, the very products it sits next to.
Each one fails at a different thing. The frame the article uses is not cloud-versus-local. It is reliability against flexibility against interface against availability against feature completeness.
The line that lands hardest in the piece is the standard the author has set: “Once you have an agent that’s available 24/7 and can actually get work done for you, you’ll never go back to a regular AI chat interface again.”
That is the bar. None of the five clears it.
A different shaped failure, same root cause
A piece on XDA from the same week looks at the question from the other side, from the developer building a local stack. Its argument is simple and very useful (source).
The bottleneck is not the GPU.
The bottleneck is not raw speed at all. It is the system around the model. The piece names three real bottlenecks, and none of them are hardware: how the workflow is designed, whether the model has access to your actual data, and how the operator structures and uses the system. A model with no connection to your life is, in the article’s words, “the real bottleneck isn’t how fast the AI thinks; it’s how much it knows about your specific world.”
Put the two pieces side by side and you get a clean diagnosis. The five most-watched agents are each missing a different thing on the surface, but the missing piece underneath is the same: an agent that genuinely lives inside the operator’s day, with their data and their tools and their habits, available all the time.
Build something that thinks fast but lives in a chat tab and you have, in the XDA author’s image, an engine that idles.
What Ostler does about it
Ostler is built on the answer both pieces are circling. A local agent that lives on the customer’s Mac, ingests their actual data, and integrates into their actual workflows.
A short list of what that means in practice.
- The Hub runs on a Mac the customer already owns, Apple Silicon M1 or newer. No fleet to manage, no second machine. Apple Silicon’s unified memory is genuinely useful here for local inference, but it is supporting cast, not the headline.
- Deep ingestion of the data that actually matters to a person’s day. iMessage, WhatsApp, email, Photos, Calendar, Safari browsing history, voice notes. The agent is not asked to act on a life it cannot see.
- A local stack tuned to that job. My own Hub, after months of use, holds roughly 148,000 embeddings of my life in Qdrant and about two million RDF triples in Oxigraph; a fresh customer install starts empty and grows from each customer’s own data. Redis runs the cache and the message bus. Ollama runs the local Qwen 3.5 9B reasoning model (about 6.6GB on disk) and nomic-embed-text for embeddings. Whisper handles speech to text. SQLCipher keeps the database encrypted at rest. A Rust agent runtime drives it all. An MkDocs compiler turns the graph into a private wiki.
- Opinionated single-machine architecture. One Mac, one install, one place where the data lives. The complexity is in the pipeline, not in the deployment.
- Default-off network paths. Anything that goes off the machine is opt-in and toggleable. The default position is that your data does not leave.
- An iOS app for when you are not at your desk. The Hub does the heavy work. The phone is the remote.
Two provisional patents have been filed on the pipeline.
There is a funding-architecture argument hiding in this too. A product that runs entirely on the customer’s own machine is awkward to build if your cap table needs you to charge per token. A product that needs to ingest a person’s whole life is awkward to build if you are also asking them to upload it to your servers. Ostler is shaped the way it is partly because it is funded the way it is.
The bar to clear
The bar Yang sets is the right one: an agent that is available all the time and that actually gets work done for you. Both articles are honest about how far the current crop is from that bar. Yang’s piece because none of the five named players gets all ten boxes. The XDA piece because raw model and raw silicon were never the problem in the first place.
We think the way to get there is the unfashionable one. Put the agent on a machine the operator already owns. Give it access to the data that actually describes their life. Wire it into the tools they already use. Keep the defaults private. Make the pipeline the product.
The Hub installer is signed, notarised, stapled, and ready. The iOS app ships through the App Store. The architecture above is what you get when you buy it.
The world DOES revolve around you.™