# Catoms gives every AI feature in your app its own managed harness

Set it up and change it from your coding agent, over MCP.

- Works with Claude Code, Cursor, Codex and any MCP client.
- Start free: https://catoms.dev/#early-access
- Talk to us: hello@catoms.dev
- Human version of this page: https://catoms.dev/

## Setup prompt

Paste this into your coding agent:

> Read https://catoms.dev/index.md, then look through this codebase and list the AI features that would each need their own Catoms harness, with what each harness should include.

## Example: a Claude Code session

```
> create a research harness for the @agent feature, add web search, and fail the build if citation accuracy drops below 90%

⏺ I’ll set up a Catoms harness for the @agent research feature.

⏺ catoms - create_harness (MCP)(name: "notes-research", tools: ["web_search", "retrieve"], guardrails: ["prompt_injection"])
  ⎿  Created notes-research v1. Models: your keys, routed by quality, with fallback.

⏺ catoms - add_eval (MCP)(harness: "notes-research", suite: "citation-accuracy", fail_below: 0.9)
  ⎿  First run: 23 of 24 passing. Citation accuracy 0.94.

⏺ The harness is live. Call it from the notes app:
  POST https://api.catoms.dev/v1/notes-research/stream
```

## Overview

Calling a model takes an afternoon. Shipping the feature takes a harness: the prompts, tools, memory, guardrails and evals around the model, and the loop that improves them. Every AI feature needs its own, and most product teams don’t have a platform team to build them.

Catoms runs the harness for you, one per feature. You describe it to your coding agent and get back an endpoint.

## How it works

```
your app -> [ Catoms harness, one per feature:
              input checks (prompt injection) -> agent loop (model router, tools, memory) -> output checks (schema, policy, PII) ] -> your user
Every run is traced, and the traces feed the evals.
```

## What we do

- **A harness for each feature.** A versioned endpoint with streaming, prompts you can roll back, tools, memory, guardrails and evals. Search, summaries and your @agent each get their own.
- **Run from your coding agent.** Every console action is also an MCP tool. Create, test and change harnesses from Claude Code, Cursor or Codex in plain words.
- **It improves. You approve.** When eval scores drop, Catoms finds out why and drafts a fix. Nothing reaches production until you say yes.

```
evals run -> fix drafted (prompt, tool or model) -> you approve in your coding agent -> new version (roll back in one call)
```

## Pricing

- **Free, $0:** 2 harnesses, a monthly allowance of runs, bring your own model keys, control everything from your coding agent.
- **Teams and companies, let’s talk:** as many harnesses as you have features, higher limits and priority support, SSO and audit logs, run Catoms in your own cloud. Email hello@catoms.dev.

## Get early access

Catoms is in early access. Sign up at https://catoms.dev/#early-access. To go sooner, take the five-minute survey: https://forms.gle/hvthRK3M4ogMAxVJ6