The product

One console. Every agent, every source, every retrieval.

This is the console we are building: who has an agent running, what it is connected to, and what it has been allowed to read. The panel below is an interactive mock-up of that interface, not a live deployment. Start an agent, connect a source, walk the graph.

Control plane agent@varrolabs.ai admin

Agent instances

NameStatePortStartedActions

Knowledge sources

Graph explorer

Name an entity and walk its connections

The Librarian

One gate in front of everything an agent may know.

Six capabilities, and every one of them is checked against the caller's clearance before it returns a single word.

read

Search memory

Meaning and keyword search combined, ranked by how close a fact sits to the question, and filtered by the caller's compartments before any result exists.

read

Look up an entity

One thing and its currently valid connections, returned as compact text with the dates each fact was true and the compartment it belongs to.

read

Walk the neighbours

Bounded traversal of what surrounds an entity. Depth limited by design, so no agent can walk its way out of its own compartment.

write

Add an episode

Takes raw text, pulls out the things and relationships inside it, removes duplicates, stamps when it was true, and retires whatever it contradicts.

read

Search the web

External search as a server-side capability, provider-swappable and rate limited, rather than an unlogged side door inside every agent.

system

The record

Every access and every change is appended: who, what, when, and what it looked like before and after. Surgical undo, full replay, and a compliance story that survives review.

Architecture

One write gate. One graph. Zero trust in the model.

Agents reason against their own model. They never hold the graph, and they never decide what they are allowed to see.

Agent terminal per-agent stream, admin panel, approvals
Control plane sign-in, quota, agent lifecycle
Agent one per person, ephemeral
Model gateway short-lived, scoped credentials
The Librarianthe gate identity, clearance, retrieval, record
Company memory episodes, entities, summaries
Nightly consolidation dedupe, supersede, snapshot

The sequence

Five steps. The third is the entire company.

Clearance is settled server-side, ahead of retrieval, so unauthorized facts never enter the process at all. Everything else is plumbing.

Step one

Someone asks a question

Plain language. No syntax, no folders, no knowing which system holds the answer.

Step two

The agent asks for context

It arrives carrying a verified identity it did not choose. It cannot assert its own role or clearance.

Step three, the product

The gate decides what exists

Clearance is applied before anything is retrieved. Facts outside it are never fetched, so no amount of clever phrasing can surface them.

Step four

A handful of facts come back

Eight to fifteen current facts with their validity dates. Never a copy of your knowledge base.

Step five

The access is recorded

Who asked, what came back, and what was withheld, written to a trail you can read without asking us.

Compartments

One company. One graph. Different truths.

Compartments are data, not hardcoded roles, so your access model is a configuration change rather than a deploy.

people

HR and People Ops

Agents that answer policy, compensation-band and onboarding questions against real internal history, without ever surfacing an individual's record to someone outside the compartment.

engineering

Engineering and IT

Incidents, runbooks, architecture decisions and the reasoning behind them, kept with their dates, so "why did we choose this in 2024" has an actual answer.

executive

Executive and Strategy

Cross-compartment visibility for the people who are meant to have it, with a full record of every retrieval, enforced by the same gate as everyone else.