Your company's privateAI workspace.
Chat, project memory, shared agents, and approved skills on your servers. Models and tools connect once, centrally, and every call is logged.
Particula⌄ParticulaA team runs on more than chat.
All of it stays on your servers.
Nine things a team reaches for daily. They all read from the same memory, and none of them phone home.
Every model, one app
200+ models through OpenRouter, or your own Ollama and vLLM. Swap the model under a team without changing how it works. Same app in the browser and natively on macOS and Windows.
Memory scoped to the project
Threads, meetings, documents, and code become one memory, bound to the team that made it. Answers cite who said it and when, and Cmd+K searches all of it by meaning.
Memory that retires itself
When a newer fact contradicts an older one, the old one is marked superseded, not deleted. Retrieval returns what is true now; the chain of what changed stays readable.
Offices inside a team
A private room for work with a smaller circle. Only its members see the threads and the memory, and someone from another team can be invited in without opening the rest up.
Agents that join the team
Build an agent once and publish it to a team. It sits there like a member: same memory, same skills, same tools, and nothing beyond what that team was granted.
A skills hub, with approvals
Anyone can write a skill; a team owner or admin approves it before it spreads. Versioned, with rollback, and set to run automatically or to ask before it applies.
One tool layer, one admin
Connect an MCP server once at org level. Credentials stay there, grants go out per team and per agent, and every tool call is logged with who, what, and the result.
Orgs, teams, roles
Multi-tenant by design: organizations, teams, offices, and four role levels. Access enforced at the database layer with row-level security, not just in the app.
Admin sees all of it
Usage and cost per model, team, and user. Agent registry, skill approvals, the full tool-call log, and an audit trail of every grant and every change.
One container,
then your identity stack.
Three steps, and the last one hands over the runbook. Nothing about running Lumen depends on us afterwards.
- 01
Single container install
One Docker image, one Helm chart. Deploys to your cluster in under an hour, including SSO and storage configuration.
- 02
Identity-native setup
SAML and OIDC SSO with SCIM provisioning out of the box. Your users sign in with existing accounts; admin roles map to your IdP groups.
- 03
Observability and handoff
Prometheus metrics, OpenTelemetry traces, and structured audit logs plug into your monitoring on day one. We stay on call for the first 90 days.
Replaced unmanaged ChatGPT use with a self-hosted workspace. Shadow AI dropped to zero; adoption did the opposite.
Two more, same
deployment story.
Every product ships as a container you host. Adding a second one reuses the stack the first one already runs on.
Thirty minutes,
then a proposal in two days.
No sales deck, and no discovery phase you pay for. We look at your stack and tell you whether it fits.
- 01
Architecture call, 30 min
With our solution lead, focused on your stack and constraints, not a feature tour.
- 02
Pilot proposal in 48 hours
Scope, timeline, integration matrix, and success metrics, in writing.
- 03
Reference call, optional
Talk to a team currently running Lumen in production before you commit.
“Shadow ChatGPT is not a policy problem, it is a tooling problem. Give people a workspace that is faster than the browser tab and remembers what the team knows, and the data stops leaving on its own.”
Sebastian Mondragon, Founder of Particula Tech