Rysh runs on Claude — why we build on the best model instead of pretending we trained one
We get asked \"so what's your model?\" a lot. Answer: Claude, on your API key. Here's why that's a feature, and where the real work is.
At some point in every conversation about rysh, someone asks: "so what's your model?"
Answer: Claude. Anthropic's. On your own API key.
And about half the time I can watch the interest drain from their face, because the startup script says I'm supposed to mumble something about "our proprietary models" or "our fine-tuned routing layer." I'm not going to. This post is the full, unhedged version of why — because I think the reasoning matters more than the fact.
The claim we will never make
Let's kill the ambiguity first, because vague model claims are the background radiation of AI marketing.
rysh does not train models. rysh does not claim to make Claude smarter. When an agent in rysh writes good code, Anthropic's researchers earned that, not us. Anyone building on a frontier model who lets you believe otherwise is borrowing credibility they didn't earn — and you should apply that test to every AI product you evaluate, ours included.
So what do we actually do? Everything the model can't.
The model is the engine. Nobody ships an engine.
A frontier model is an extraordinary component and a useless product. On its own it can't touch your filesystem, can't run your tests, can't be interrupted safely, can't ask permission before something destructive, can't answer your website chat, and can't show your team what it's doing.
The distance between "brilliant text predictor" and "agent doing real work in your company" is the actual product surface:
- Execution. 46+ built-in tools — files, git, shell, search, test running, web — plus every CLI you already own, wired to a real PTY-backed terminal.
- Governance. Every tool call visible in a pane. Coloured diffs before an edit lands. Mutating commands gated behind explicit approval; a safe-by-default allowlist so read-only work flows freely.
- Multiplicity. Not one assistant — a workspace of named agents (
@code-reviewer,@ops-bot) spawned from markdown files, coordinating across panes. - Surfaces. The same agent definition answering in a terminal pane, driving a browser, or living on email and website chat and Slack as a humanoid (WhatsApp/phone: roadmap).
- Collaboration. Live pane sharing across machines, view or control, per share.
- Operations. Headless control (
rysh send), session persistence, detach/attach — agent infrastructure you can script.
Notice that not one item on that list is "be smarter than Claude." That's the point. We picked the layer where a small team can genuinely be the best in the world, and it isn't pretraining.
Why Claude specifically
Honest reasoning, no scorecard theater: agentic work is not chat. What matters is sustained tool-use loops, respecting constraints over long horizons ("show the diff first," "never touch prod"), and knowing when to stop and ask. In our building and daily dogfooding, Claude has been the model we trust most for exactly that shape of work — which is why it's the default engine today.
And because incentives matter more than compliments: rysh's architecture keeps the provider layer pluggable. We stay on Claude because it keeps earning the seat, not because we're stuck. That's the difference between a choice and a dependency.
Your key, and why that's the quiet headline
Here's the part with real consequences for a company: rysh runs on your Anthropic API key, on your machines. rysh is self-hostable (self-hostable, to be clear — not open source today).
Follow the data path. Your prompts, your code, your customer conversations go from your machine to Anthropic under your agreement with them. Not through our servers. Not into a mystery middle layer that "may use data to improve services." When an agent in rysh reads your repo, the parties involved are you and your model provider. We designed ourselves out of that loop on purpose.
It also means no markup arbitrage. Products that resell model access by the seat are incentivized to route you to cheaper models quietly. With BYO key, you pay Anthropic for the model and us for the platform — two transparent bills, aligned incentives.
One more layer on the trust story: when agents handle credentials, SecretNAT (##snat) swaps real secret values for stable tokens before anything reaches the provider. We verified this on the wire — the provider receives the token, never the value. Building on Claude doesn't mean shipping Claude your secrets.
"So you're just a wrapper"
The dismissal writes itself, so let's face it head-on.
Every product on a frontier model is "a wrapper" the way every application is "a wrapper around an operating system." The question is never whether there's a model underneath — it's whether the layer on top does something hard.
Try to build rysh's layer this weekend: a real terminal multiplexer with a proper terminal emulator; an actor-model runtime where every pane is an addressable participant over an embedded message bus; approval gates at the tool layer; separable output streams per pane; live cross-machine pane sharing; markdown-defined agents deployable to email, web chat, and Slack; headless session control. That's years of systems engineering, and none of it appears in the next model release — any model release makes it more valuable, because better engines deserve better vehicles.
Model capability is the rising tide. We're deliberately building the boat, not competing with the ocean.
The test to apply to everyone (including us)
When you evaluate any AI product, ask: "what's yours?" If the answer leans on the model's brilliance, you're buying a markup. If the answer is concrete about execution, governance, surfaces, and ownership — things that exist around the model — you're buying engineering.
Our answer is one binary away:
brew install rysh-ai/rysh/rysh
rysh
Bring your own key. Watch every tool call as it happens. That transparency is the pitch.
We're recruiting design partners. If you want an agent platform that's honest about where the intelligence comes from — and serious about everything around it — come shape the roadmap: rysh.ai/design-partner
Related: Meet rysh.ai · Rysh vs building it yourself