The technology behind private AI that does the work
BrainOutput is one chat interface running on your own hardware or in a private cloud. It works with any model, routes every task to the right one, and keeps a durable memory of the work. This page explains how: the routing, the stack, the memory, and the audit trail.
Model routing by confidentiality
Every request is classed as sensitive or not sensitive. The class — not the model vendor — decides where the request goes.
Sensitive stays private
Client files, contracts, accounts, health data, guests' data: these requests are answered by a private model, on your machine or in your private cloud. They never leave.
Non-sensitive goes where it makes sense
A public text, a generic draft: these may go to a cheaper or stronger cloud model — but only when your policy allows it.
You set the policy
The routing rules are yours, per team and per use case. On a new installation everything stays private; a cloud model is used only if you enable it.
Your sensitive work never leaves; the rest goes wherever is cheapest or best — only when you say so.
The stack, component by component
Each component does one job and can be replaced without rebuilding the rest. Nothing in the stack ties you to a single AI vendor.
- DevicesOpenClawyour PCs and phones
- InterfaceBrainOutput chatbuilt on LibreChat
- OrchestrationOrchestratorPaperclip · plans, tracksAgentsOpenCode · isolated workspaces
- Memoryper user, in the chatper project, on the boardknowledge, in files
- Model gatewayModel gatewaystable aliases + confidentiality routing
- ModelsLocal modelsopen-weight · vLLMPrivate clouddedicated inferenceCloud models (BYOK)only if your policy allows
LibreChat — the interface
One chat interface for the whole team: conversations, agents, files, per-user accounts.
Your team learns one tool. Every capability — the platform and all its add-ons — is used from that same chat, not from another portal.
Paperclip orchestrator — the company spine
Plans work, breaks it into tasks, assigns them to agents, and tracks everything on a board.
Long, real projects run to completion — not one-shot answers. Every task, decision and comment is recorded, so the work is auditable by design.
OpenCode agents — the executors
AI agents that execute tasks: writing code, analysing documents, preparing deliverables.
Each agent works in an isolated workspace on its own branch. Work is reviewed and merged like any professional engineering process — no agent touches production directly.
OpenClaw — the devices
Connects PCs and phones to the platform.
The work follows you across your devices instead of living in one browser tab.
Model gateway and aliases — the switchboard
A single gateway addresses every model by a stable alias.
Models are swapped — for cost, speed or a better release — without changing a single workflow. The alias stays; the model behind it is your choice.
Local vLLM inference — the private engine
Open-weight models served locally on your hardware through vLLM.
Sensitive work is answered at home, with near-unlimited use at a fixed cost — no per-token meter on your confidential data.
BYOK — your keys, your cloud models
Bring your own keys for the cloud models your policy allows.
You keep your existing provider relationships and contracts. We never resell tokens and never stand between you and your model vendors.
Memory management
An AI that forgets everything between sessions cannot run real work. BrainOutput keeps memory in three explicit layers — you always know what is remembered and where.
Conversation memory, per user
Each person has their own memory in the chat: preferences, context, ongoing work. It is per account, not shared by default.
Project memory, on the board
Durable work lives as tasks, documents and decisions on the orchestrator's board. A project survives any single conversation — and any single model.
Knowledge, in files
Lasting knowledge is written to versioned files, not trapped in a model's context window. It can be searched, reviewed, corrected and audited like any company document.
Bounded by design
Agents read only what a task needs, and buffers and lists are capped explicitly. Memory that grows without bound is a leak; ours has edges you can inspect.
Security and audit
Private AI is not a promise on a landing page. It is a set of mechanisms you can check.
- Data residency you choose. The platform runs on your hardware or in your private cloud. Sensitive requests are answered there; the routing policy — not a vendor's terms of service — decides what may leave.
- A complete audit trail. Every task the orchestrator runs carries its history: who asked, which agent executed, what changed, which branch carried it. You can reconstruct any piece of work after the fact.
- Review where it matters. Risky changes — install, secrets, authentication, infrastructure, personal data — go through a review stage before they merge. Routine work closes on the executor's evidence.
- Scoped, revocable access. Agents act with scoped credentials, inside isolated workspaces, under budget and approval gates with kill switches. No agent holds standing access to your production systems.
- No lock-in. The stack is LLM-agnostic and every component is replaceable. Leaving any single vendor — including us — never means rebuilding your workflows.
See the stack running on real work
We will show you the routing, the memory and the audit trail on a live system — with demo data, never a client's.