ALL AGENTS NOMINAL1,243 TICKETS TODAY94% AUTO38S MHT
CHANGELOG · v1.4BERLIN / SF · MAY ’26
the operator OS, in five layers.

Build the agent. Trust the receipts.

Lumin is the canvas, the runtime, and the audit trail for operator-facing AI. Define agents visually, plug them into your channels and systems of record, ship them in days, and trace every action they take.

01
Channels
02
Builder
03
Runtime
04
Receipts
05
Trust
the Builder, on a real agent.
app.luminflows.io / builder / freight-intake / v3.4.1
SAVED · 0:08 AGO
NODES
@Input
Parse
?Plan
Tool
Decide
Write
Log
INSPECTOR · PLAN
Plan · book
MODEL
claude-sonnet-4-5
PROMPT
Given the parsed load, propose carrier + price within margin band.
EVAL
94%↑ 2.1
FIVE LAYERS

The whole product, top to bottom.

Channels bring the work in. The Builder defines what the agents do. The Runtime picks the right model. Receipts explain every action. Trust is the floor, not a feature. Each layer earns the next.
↓ each one drills in further.
01
Channels
What comes in. What goes out.

Lumin meets your customers where they are and posts where you keep score. 100+ inbound channels, 100+ outbound systems, REST + webhook for the rest. Same agent, every channel.

Email · phone · WhatsApp · SMS · Teams · web · EDI
TMS · ERP · CRM · helpdesk · DB · SAP · Stripe
REST + webhook for anything else
02
Builder
Visual canvas. Production code.

A node-graph editor for your team. Drop nodes — Parse, Plan, Tool, Decide, Write, Log — wire them up, set guardrails, ship. No code; production-grade execution underneath.

Node-graph with live preview
Model + prompt + tools per step
Eval suite, rollback, branch + merge
03
Runtime
The brain that picks the model.

Frontier reasoning where it matters, the smallest model that hits the bar everywhere else. Cost + latency + accuracy budgeted per step. Hot-swap any model, never break the contract.

Claude · GPT · open-source side-by-side
Per-step model + budget
Caching, retries, fallbacks built in
04
Receipts
Every action explains itself.

For every agent action, a receipt: what came in, what rule fired, what the agent did, what the human needs to look at. Exportable, signed, audit-ready by default — not as a feature flag.

JSON + PDF + Slack/email card
Signed, time-stamped, immutable
Drill-in to every model call
05
Trust
Auditors didn’t have to argue.

SOC 2 Type II, GDPR, DPA, EU + US tenants. Dedicated cloud + VPC on Enterprise. We don’t train on your data. Audit log shows every byte the agent touched.

SOC 2 Type II · GDPR · DPA
Dedicated cloud · private VPC
SSO · SCIM · role + audit
ANATOMY · ONE AGENT

Freight Intake, laid open.

nothing is hidden.
INPUT
Email
ops@northgate.com
PARSE
Lane
Berlin → Rotterdam
PLAN
Carriers
3 candidates · 1 picked
GUARD
Margin
≥ 12% · pass
TOOL
TMS
POST · TCK-1742
WRITE
Reply
to ops@northgate.com
LOG
Receipt
JSON + PDF · signed
0.1s
0.2s
0.4s
<0.1s
0.2s
0.1s
<0.1s
THE RECEIPT

One ticket. Every byte.

Receipts are the heart of the product. Each one captures the inputs, every model call, every tool call, every rule and its outcome — signed and immutable. You don’t audit the AI; you read the receipt.

JSONmachine-readable · webhook-ready
PDFhuman-readable · sealed + signed
Slacksummary card with deep link
Emailfor the customer · in your tone
RECEIPT · TCK-1742
SIGNED · 17 MAY 14:32:08
INPUTEmail · ops@northgate.com · 17 May 14:31:55
AGENTFreight Intake · v3.4.1
MODELSsonnet-4-5 (parse), haiku-4-5 (write)
TOOL CALLSTMS · POST /loads · 200 · 188 ms
GUARDRAILmargin ≥ 12% · pass (13.4%)
OUTPUTReply sent · TCK-1742 created
COST€0.041
SHA-256 · 9f3a…b27c✓ VERIFIED
INTEGRATIONS · 100+

Connects to what you already run.

+ REST + webhook for the rest.
Email
Phone
WhatsApp
SMS
Teams
Slack
Live chat
Web form
Webhooks
EDI
PDF · fax
SFTP
S3
GDrive
Dropbox
Box
TMS
ERP
CRM
Helpdesk
Database
SAP
Oracle
NetSuite
Stripe
QuickBooks
Xero
Salesforce
HubSpot
Zendesk
Intercom
Linear
+ 70 MORE · SEE THE FULL LIST →
SECURITY + TRUST

Built for auditors first.

policy + receipts in one place.
SOC 2
Type II
GDPR
EU residency
DPA
signable
SSO + SCIM
Okta · Azure
VPC
dedicated
KMS
BYO keys
NO TRAIN
ever
AUDIT
every byte
BY THE NUMBERS

What the platform does, at scale.

18M
tool calls / mo
across customers
1.1s
median agent run
across all 5 layers
99.97%
uptime
trailing 90 days
€0.04
median cost / ticket
p50, May 2026
FAQ

Product, more deeply.

The questions an engineering leader actually asks before signing.

01
What models can I use?

Any frontier model (Claude, GPT, Gemini) and most open-source ones via your inference provider. Pick per step, swap any time. Lumin never locks you in.

02
Can I edit an agent that’s running?

Yes — branch, edit in the Builder, run eval, merge. Production stays on the previous version until you promote.

03
How do I bring my own tools?

REST endpoints, OpenAPI specs, and webhooks are first-class — Lumin generates a tool node from any spec. For deeper hooks, write a custom node in TypeScript.

04
What about latency?

Median agent run is 1.1s. p95 around 3.4s. Each step has its own budget; the runtime chooses caching + fallbacks automatically.

05
How does this differ from a workflow tool?

Workflows assume the path. Agents assume the goal. Lumin gives you both: deterministic nodes where you want certainty, model-driven nodes where you want judgment.

06
Do you support on-prem?

On Enterprise — dedicated cloud + private VPC, with optional BYO KMS. Pure-on-prem is not officially supported; talk to us.

READY WHEN YOU ARE

See the product. Touch the receipts.

we’ll bring your sample tickets.

15-minute call to walk through the Builder, the Runtime, and a real Receipt on your data.

Book a demoor read the docs