Tell your coding agent what the feature should do. Catoms sets up the prompts, tools, memory, guardrails and evals around the model, hands you an endpoint, and keeps it working.
> create a research harness for notes, add web search, and fail the build if citation accuracy drops below 90%
catoms.create_harness (MCP)
name notes-research
models your keys, routed by quality, with fallback
tools web_search, retrieve
guardrails prompt-injection screen on input and fetched pages
catoms.add_eval (MCP)
suite citation-accuracy, fail below 0.90
runs on every config change
Your harness is live. First eval run: 23 of 24 passing, citation accuracy 0.94.
POST https://api.catoms.dev/v1/notes-research/stream
Say you make a notes app. Your users want to type @agent, ask a question, and watch a short research report appear in their note, with sources. Calling a model takes an afternoon. You have a demo by Friday, and it's good.
The demo makes things up when sources are thin. A pasted web page tells it to ignore its instructions. A provider goes down during your launch week. Someone asks whether last Tuesday's prompt change made answers worse, and nobody can say.
That work has a name: the harness. It's the prompts, tools, memory, guardrails and evals around the model, and the loop that improves them. The harness, not the model, decides whether an AI feature ships. You're not the only one finding this out.
And you need a harness for every feature: search, summaries, autocomplete, the @agent. Most product teams don't have an AI platform team to build them. Two engineers doing it full time cost about $420,000 a year before infrastructure, at the average US pay for an AI engineer (Thirsty Sprout).
You describe a harness for one feature and get back an endpoint. Everything between your app's request and a response you can trust is ours to run.
Every run is traced, and traces feed the evals. Each harness runs in its own microVM, so a bad tool call can't reach anyone else's.
Everything you can do in the Catoms console is also an MCP tool. Stay in Claude Code, Cursor or Codex and say what you want in plain words. Your agent does the clicking.
> the research harness cites blogs too often. prefer primary sources and show me the eval diff before you ship it
Drafted version 8 of notes-research: ranks primary sources first, drops pages without an author or date.
citation-accuracy 0.94 → 0.97 on 24 cases. Nothing else got worse.
Waiting for your approval. Nothing ships until you say so.
Most managed harnesses launched this year are tied to one model family or one cloud. But over three quarters of teams already use more than one model provider (LangChain). Catoms routes each step to the model you pick for cost, speed or quality, falls back when a provider fails, and runs on your own keys.
Your harnesses are plain files in an open spec. Export them and run them somewhere else any time. If you stay, we'd like it to be because Catoms works, not because leaving is hard.
When scores drop or a new kind of failure starts showing up, Catoms works out why and drafts a fix: a prompt change, a tool, a different model for one step. It shows you the eval results and waits. Nothing changes in production without your approval.
$0
Let's talk
We're opening Catoms to a small group of product teams first, so we can help each one set up properly. Tell us a little about you and what you want to build.
We'll write to you when your spot opens. We're letting people in a few teams at a time so we can help each one set up properly.
One favor while you wait: our survey takes about five minutes and asks how you build AI features today. Your answers decide what we build first, and people who answer get in sooner.