Skip to main content
Every HQ conversation runs on a real, hardware-isolated machine, and that machine runs qOS — the operating system we built for agents. It is written in Rust, top to bottom. We didn’t staple it together from other people’s APIs; we built the data layer ourselves, so agents never wait on someone else’s infrastructure. For where this fits in the wider picture — the roster of specialists and the machine behind each conversation — see How HQ works.

​
Why build the layer ourselves

The systems below run in production for large-scale ingestion. Building them ourselves, in Rust, is a choice about performance and robustness: an agent reaches search, cache, storage, a graph, email, a browser, documents, and extraction without leaving the platform, and the model layer and memory sit right next to them, sharing the same hardware, isolation, and audit trail.

​
The data layer

Search

Hybrid keyword and semantic search across everything, served from memory. Typesense-compatible, rebuilt in Rust.

Cache

Microsecond-class shared memory — a multi-threaded, Redis-class in-memory store with a same-machine shared-memory path.

Storage

Object storage for every artifact — an S3-compatible, masterless store with self-healing and erasure coding.

Graph

Query connected data — an in-memory graph database speaking openCypher and ISO GQL over Neo4j-compatible Bolt.

​
Content & I/O

Email

Send and receive real email — a two-way mail platform that authenticates and classifies inbound, with the full DKIM, SPF, DMARC, and ARC stack.

Browser

Drive a real web browser — a custom browser built in Rust that behaves like a person on the live web. See Web & browser control.

Documents

Read and parse PDFs — a native-quality document engine with text extraction, forms, redaction, and signing.

Extraction

Pull data out of any file — one Rust library for image metadata and PDF text, images, and structure.

​
Runtime

Inference

Frontier models, and our own — multi-model routing across providers, plus open models on GPUs we run. See Models & inference.

Memory

Remembers and learns from work — a layered memory system that consolidates between sessions, with user-facing governance.

Skills

Composable, per-agent skills — enable, disable, or tailor them per agent; an agent can install what else it needs.

Integrations

Connect the tools you use — MCP-native, with local CLIs first-class, all in one safe environment.
Each agent reaches a broad range of tools through MCP and locally installed CLIs and skills, all in one sandbox that behaves like a normal Linux server, fully the agent’s.