catoms/ˈkæt.ɒmz/noun, short for claytronic atoms
Tiny modules of programmable matter that join to form any shape. We build harnesses the same way: the same parts, assembled differently for every feature.
A managed harness for every AI feature in your app.
Describe a feature to your coding agent. Catoms builds the harness around the model and gives you one endpoint to call.
1
Install Catoms in your coding agent
/plugin marketplace add catoms-dev/plugins
/plugin install catoms@catoms-devcodex mcp add catoms --url https://mcp.catoms.dev/mcpopencode mcp add catoms --url https://mcp.catoms.dev/mcpnpx @catoms/cli init --client cursorhttps://mcp.catoms.dev/mcpNo coding agent? Paste this into any assistant that supports MCP:
Set up Catoms for me. Follow the steps at https://catoms.dev/install.md, then ask me which AI feature in my app should get a harness first.
2
Ask for a harness
Create a Catoms harness for our search feature.
A model on its own isn’t a feature.
Every AI feature needs its own prompts, tools, memory, guardrails and evals. Most teams build that by hand, one feature at a time, and maintain it forever.
This year, the harness became the unit of work.
Model makers and clouds now sell managed harnesses, but each is tied to their models and you can’t change it. Meanwhile, your coding agent can set up infrastructure just by asking.
Catoms builds the shells around the model.
Standard, tested parts, our catoms, assembled into a different harness for each feature. Each harness runs in its own machine, is versioned, and exports as plain files.
You describe the feature. Catoms builds the harness.
Say what it should do, check the plan Catoms drafts, add your files and keys, and call the endpoint.
Every request takes the same path.
In through auth and guardrails, through context and the router to a model, and back out through guardrails to be scored.
The router picks the model for each request.
It sends each kind of query to the right provider on quality, cost and speed, and falls back when one fails.
Tools come from a shared registry.
Use ready-made tools, connect your own APIs and MCP servers, or publish a tool for other teams.
Evals are written for each feature.
When scores drop, Catoms drafts a fix. Nothing ships until you approve it.
Any model, any stack, run from your agent.
Each harness is an HTTPS endpoint, and everything in the console also works over MCP.
One harness for each feature, from the same parts.
Built for product teams shipping several AI features without an AI platform team. No framework to adopt, and no single model or cloud to commit to.