Why I built Token Stanza

The origin story. A personal chat app for a model I wanted to keep, with a prompt I can read, memory and files I own, and a bill that only counts what I use.

Token Stanza is a chat app for Claude that I built for myself and a handful of people. This post is about why it exists. It is the first of a series. The posts that follow are the design decisions behind it, one at a time. The short version: I wanted to keep a model I had learned to work with, see exactly what it was being sent, and own what it remembered.

Where it started

Every few months claude.ai and ChatGPT add a new model, and a while after that an older one quietly leaves the list. For me it was Claude Sonnet 4.5 on May 18, 2026. The new model is usually better on the benchmarks. That is not the point. By the time a model disappears, I have spent months learning how it writes. I know where it is careful and where it is sloppy, and how it responds to the way I phrase things. I know that when I ask it to “consider alternatives,” it will actually list them rather than defending its first answer. I know it writes clear commit messages without being prompted and that it tends to over-explain obvious points in technical writing unless I tell it not to. All of that is gone the morning the menu changes. Just as you get used to one, they want you on the next.

On May 18, 2026, Claude Sonnet 4.5 was removed from claude.ai. It was still reachable from Claude Code, so I used it there for a while. But Claude Code is a tool for working on code, not a place to have a conversation. The prompt is the tool’s, the memory is the tool’s, and a chat that is not about code is a poor fit for it. The model I wanted was still being served by three platforms: Amazon Bedrock, the Anthropic API and Google Cloud. What was missing was a plain chat window in front of it. One that would not disappear the next time someone else changed the menu, and that I could open to a few other people who wanted the same thing.

That was the moment I decided to build rather than wait for the next menu change. The first commit landed in August 2026, and the project was called Sonnet45Chat, after the one model it existed to keep.

What a personal app gives you

Building my own app meant I could have four things that mattered to me, in the order I cared about them.

A readable prompt. The View prompt panel shows the built-in instructions, the custom prompt, the memory snapshot and the tools, with token counts.

The Prompt panel: the whole prompt's token count, the model and thinking level, then the system prompt sections (built-in, custom prompt, memory snapshot, conversation tools) with a token estimate each
The View prompt panel. Every section the model is sent, with its size, and a line saying whether it matches what the last turn actually sent.

Memory and files I own. Every note the model keeps and every file it writes is readable, editable and downloadable from the Memories and Files pages. The whole account exports as a zip from Settings.

A model I pick and keep. A conversation is pinned to one provider and one model for its life. The landing page says the model stays available for as long as the platforms that host it keep serving it. By the time you read this, two of the three have published retirement dates for Claude Sonnet 4.5, and Google has said nothing. A later post covers what “retiring” actually means, because it means three different things on the three platforms and the useful life is longer than the announcement suggested.

Pay-as-you-go. A prepaid balance, each reply charged for what it used, and a Usage page that says where the balance went. I wanted to know exactly what I was paying for. A flat monthly tier never told me whether it was too much or too little for how I actually used it.

The Usage page: the current balance, tiles for messages, estimated cost, tool calls and each token bucket, and a per-day table of tokens by kind with a cost column
The Usage page. Each day's tokens by kind, priced at the provider's list rates, including the 10% US-routing premium where it applies, next to the balance they came out of.

The third one is the reason the other three exist. Once you are paying a provider directly for a model, there is no subscription tier deciding which models you see. There is no product team deciding when a model has had its turn. The model goes away when the platform retires it, which is a published date with notice, and not before. The other three fell out of having the app in my own hands. If I am building the prompt, I can show it. If I am storing the memory, I can hand it back. If I am paying per token, I can show the bill.

What it is today

Three months in, it is one container on AWS Fargate serving a Vue client and an ASP.NET Core API over Claude on three cloud providers. Streaming replies with prompt caching. A per-user memory tree and file library the model can read and write. Web search and page reading, background summaries, keep-warm pings that hold the prompt cache, and usage reporting. The features section of the landing page has the user’s view of that list. The rest of the series is the design decisions behind each item.

A conversation in Token Stanza: the conversation list on the left, a reply with a collapsed thinking block and a per-turn usage line, and the composer with its thinking, provider and keep-warm controls, a context percentage and a cache countdown
A conversation. Under each reply is what it cost in tokens. Under the composer is how full the context is and how long the prompt cache stays warm.

The product got its name later. The repository, the projects, the auth cookie, the browser storage keys and the AWS resources are all still called Sonnet45Chat. They will stay that way. Those are identifiers, and several of them are stored data. Renaming the cookie would sign everyone out. There is a dedicated post about the rename and the line it drew.

Who it is for

Anyone who wants to keep talking to a model after the big chat products have moved on from it. Registration is whitelisted and I review every request by hand. That is a feature, and a later post explains what a tiny user table lets a solo developer skip. If that is you, request access is the way in.