# Lumni: Catch AI agent failures that look like success

Detect duplicate actions, conflicting updates, and false success claims across your business systems. Start in watch-only mode in minutes.

- Book a 15-min Hardening Call: https://cal.com/anuragvishwa/30min
- See it work in a minute: `npx lumniverse try`
- Start watch-only free: https://app.lumniverse.com/sign-in?signup=1 (never blocks on day 1)
- Live demo of the console: https://app.lumniverse.com/today?data=mock
- Docs for agents: https://lumniverse.com/llms.txt (everything in one file: https://lumniverse.com/llms-full.txt)

## Three reasons it's safe to try

- **Watches first.** Nothing is blocked until you turn it on, per action.
- **Plain rules decide.** Deterministic checks you can read, not a model.
- **Open source.** Self-host it free. Your data never leaves.

Works with LangGraph, OpenAI Agents, Temporal, MCP and Claude Code. SDKs in TypeScript, Python and Go.

## Your agents, with and without Lumni

| Situation | Without Lumni | With Lumni |
| --- | --- | --- |
| A retry after a timeout refunds again. | Paid twice | Lets one through |
| You fix a record; an agent overwrites it. | Fix lost | You always win |
| The tool said OK. The app has nothing. | Silent failure | Proves it in the app |
| One task fans out to forty agents. | Runaway spend | Slows it at your limit |
| Charged the card; the deal never closed. | Unnoticed | Flags it as an incident |

## How to start

1. **See it work.** `npx lumniverse try`: sample agents make real mistakes in a sample workspace of your own, and you watch Lumni catch them.
2. **Connect your agent.** `npx lumniverse init` in its folder. Lumni starts by only watching: you see what it would stop, and nothing is blocked.
3. **Protect.** When the record shows it only stops real problems, turn it on per action (Settings → Protect). Approvals land in the console.

Or let your coding agent add it: `npx skills add lumnihq/lumni`, then ask "Add Lumni to the refund step." Works in Claude Code, Cursor, Codex, Copilot and more.

## What Lumni does

- **Stops the repeats.** Two agents, one refund: paid once. When a retry or a second agent tries the same thing, Lumni lets the first through and stops the copy, and shows what each repeat would have cost.
- **Shows every action.** One picture of your whole fleet: what went straight through, what looked again, what waited its turn, and what Lumni stopped.
- **Takes turns.** People go first; agents wait. When an agent and a person reach for the same record, Lumni makes them take turns, and makes an agent re-read anything that changed under it.
- **Asks you.** Risky actions, like a refund over your limit, wait for a yes. Approve or reject in one click. Slack and email alerts.
- **Proves it.** Lumni reads the result back from the app itself, not the agent's word. Half-finished work becomes an incident with what to do next.

## The core call

```ts
const d = await lumni.checkAction({
  actor_id: 'support-agent', work_id: ticket.id,
  operation: 'stripe.refund',
  effect_key: `refund:${order.id}`,
});
if (shouldExecute(d)) {
  await stripe.refunds.create({ charge });
}
```

Decisions: ALLOW, DEDUPLICATE, WAIT, REPLAN, REQUIRE_APPROVAL or REJECT.

## After 48 hours in watch-only

Watch-only never blocks. It records every action it would have stopped, and after two days you get a risk report, for example: 3 duplicate payment retries caught ($1,840 saved) · 6 human-rep overwrites it would have blocked · 2 half-finished ERP writes flagged.

Built for the systems you can't afford to break: Stripe Refunds · Shopify Orders · NetSuite / QuickBooks · Zendesk / Gorgias · Salesforce / HubSpot.

## Pricing

| Plan | Price | Includes |
| --- | --- | --- |
| Watch-only | $0/mo | 1,000 events/mo included. Watch-only (never blocks), SDKs + MCP + coding-agent skill. |
| Production | $299/mo | + $0.003 / event up to 1M, then $0.002 / event. Stops repeats, stale writes and collisions, Slack approvals, read-back verification. |
| Enterprise | Custom | Pricing tailored to your team. We wire Lumni into your stack in 48 hours: hands-on integration (LangGraph, Temporal, MCP), 48-hour Shadow-Mode Risk Report, custom read-back verifiers (Stripe, Shopify, ERP, CRM), a dedicated Slack channel. $0 if we don't catch a real bug in week 1. |

One event is one action an agent asks Lumni about before doing it: a refund, an email, a CRM write. Reads, dashboards and people using the console are free.

## Open source

The server and console are FSL-1.1 and become Apache-2.0 two years after each release. The SDKs, the CLI, the skill and the examples are Apache-2.0 today. Code: https://github.com/lumnihq/lumni

## Frequently asked questions

### What is Lumni?
Lumni is a check your AI agents call before a consequential action, like a refund, an email or a CRM write. It stops repeats, stale writes and collisions with people, asks you about risky actions, and reads the result back from the app afterwards.

### Who is Lumni for?
Ops, support and engineering teams that let AI agents act in real systems such as Stripe, Zendesk, Salesforce or a database, and need each action to happen once, the way they meant.

### Will Lumni break my agents on day one?
No. New workspaces only watch: Lumni records what it would have stopped and lets everything through until you turn protection on, per action.

### What's the difference between try and init?
`npx lumniverse try` runs sample agents in a sample workspace so you can see Lumni catch real mistakes. `npx lumniverse init` connects your own agent to your real workspace.

### What counts as an event?
One call from an agent asking about an action it's about to take. Reads, dashboards and people using the console are free.

### Does an AI decide what gets stopped?
No. Whether an action runs is decided by plain, deterministic rules: the same effect twice, data that changed, someone else on the record, or a rule you set.

### Is my data safe?
Each workspace is isolated, keys are hashed, and we never sell your data or train on it. Privacy policy: https://app.lumniverse.com/privacy

### How do I disconnect a machine?
Settings → Machines and keys. Each connected machine has a Disconnect button; it's refused from its next call.

---

Lumni is made by Chainslice Inc. Contact: privacy@lumniverse.com
