# Lumni > Lumni is the coordination layer for AI agents. Before an agent takes a consequential action (a refund, an email, a CRM write, a deploy), it asks Lumni; Lumni checks the action against what is true now and what other agents and people are doing, and answers ALLOW, DEDUPLICATE, WAIT, REPLAN, REQUIRE_APPROVAL or REJECT. Afterwards Lumni verifies the result in the target app instead of trusting the tool call. No LLM decides; the hard path is deterministic. Key facts: - Cloud: https://app.lumniverse.com (Watch-only $0/mo with 1,000 events/mo included; Production $299/mo + $0.003 per event up to 1M, then $0.002 per event; Enterprise: custom pricing, and the team wires Lumni in within 48 hours). Self-host with Docker Compose. Server and console are FSL-1.1 (Apache-2.0 after two years); SDKs, CLI, skills and examples are Apache-2.0. - See it work: `npx lumniverse try` runs sample agents that make real mistakes in a separate sample workspace while Lumni catches them (it creates an account if needed). Connect a real agent: `npx lumniverse init` in its folder (writes `LUMNI_URL` and `LUMNI_API_KEY`). - SDKs: TypeScript `@lumniverse/sdk` (Node 22+), Python `lumni-sdk` (3.10+), Go `github.com/lumnihq/lumni/sdk/go`. Adapters for LangGraph, OpenAI Agents, Temporal, Restate, and an MCP gateway. - Core call: `checkAction({ actor_id, work_id, operation, effect_key, resources, arguments, metadata })`, then act only if `shouldExecute(decision)`, then `startAction` and `reportResult`. - `effect_key` names the real-world effect (e.g. `refund:`): the same key means the same effect, which happens once. - New workspaces run in shadow mode: decisions come back as `WOULD_*` and nothing is blocked until enforcement is turned on. ## Docs - [Lumni skill for coding agents](https://raw.githubusercontent.com/lumnihq/lumni/main/skills/lumni/SKILL.md): how to add Lumni to an agent, field by field, and how to handle each decision - [README](https://raw.githubusercontent.com/lumnihq/lumni/main/README.md): what Lumni is, getting started, licensing - [Getting started](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/getting-started.md): one complete workflow, end to end - [SDKs](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/sdk.md): Go, TypeScript and Python clients and durable runtimes - [TypeScript SDK](https://raw.githubusercontent.com/lumnihq/lumni/main/sdk/typescript/README.md) and [Python SDK](https://raw.githubusercontent.com/lumnihq/lumni/main/sdk/python/README.md) - [API](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/api.md): authentication, admission and decisions; the full contract is [api/openapi.yaml](https://raw.githubusercontent.com/lumnihq/lumni/main/api/openapi.yaml) - [Integrations](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/integrations.md): MCP gateway, OpenTelemetry, GitHub, webhooks ## Running Lumni - [Self-hosting](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/self-hosting.md): Docker Compose, HTTPS, updates - [Configuration](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/configuration.md) - [How coordination works](https://raw.githubusercontent.com/lumnihq/lumni/main/docs/coordination.md) ## Optional - [Examples](https://github.com/lumnihq/lumni/tree/main/examples): minimal agents in TypeScript, Python and Go, Temporal and Restate workers, and `agent-fleet` (agents that make realistic mistakes) - [Security](https://raw.githubusercontent.com/lumnihq/lumni/main/SECURITY.md) - [Privacy policy](https://app.lumniverse.com/privacy) and [terms](https://app.lumniverse.com/terms) - [Everything above in one file](https://lumniverse.com/llms-full.txt)