@mention a specialist and it researches, decides, and does the work, then reports back. This page explains the machinery behind that, and where each piece surfaces in the API.
A roster of specialists
You don’t get one generic assistant — you assemble specialists, each with its own model, skills, integrations, and tone, reachable by an@handle. The right specialist picks up a thread and stays with it from request to done.
→ Configure them with the Agents API (agents:read / agents:write).
A real machine behind every conversation
The difference between chatting and doing is a computer. Every conversation runs on its own private, hardware-isolated machine — the same class of isolation behind modern serverless platforms. It boots in a fraction of a second and resumes exactly where it left off, even days later. Because it’s a real machine, a specialist can run a script, crunch a file, drive a browser, build a small tool, and install what it needs — instead of only telling you how. → Drive that machine through conversations and turns; deliverables come back as artifacts and live surfaces.qOS — the agentic operating system
That machine runs qOS, the operating system we built for agents — in Rust, top to bottom. We didn’t staple it together from other people’s APIs; we built the data layer ourselves:Search · Cache · Storage · Graph
Hybrid search across everything, microsecond shared memory, object storage for every artifact, and a queryable graph of connected data.
Email · Browser · Documents · Extraction
A real inbox that sends and receives, a real web browser, PDF parsing, and structured extraction from any file.
Inference · Memory
Frontier-model routing, plus memory that remembers and learns from work.
Skills · Integrations
Composable, per-agent skills and connectors to the tools you already use.
The right model for each job
Decision intelligence needs more than one model. HQ routes every call across the frontier — Anthropic, OpenAI, xAI, and Cerebras — and sends each task to the model that wins on quality, latency, or cost. The high-volume, data-sensitive work — the embeddings and reranking behind memory and search — runs on open models on our own GPUs, so your most sensitive text never leaves our infrastructure to be turned into numbers. → See what’s available via List models, and select one per agent.Research you can act on
HQ specialists browse the live public web the way a person does — from the region you choose, reaching sources and logged-in tools an ordinary assistant can’t — and hand back answers with their sources, not a confident guess. Agents can also drive the user’s own browser. → Browser control API.Memory that compounds
Specialists remember and consolidate between sessions, and arrive knowing your workspace’s channels and people. You can always see what a specialist knows, and correct, delete, or freeze it. → Memory API (memory:read / memory:write).
Trust is the foundation, not a setting
Anything that can act needs a trust model that holds. Every action a specialist takes is bound to the person who asked and the permission they were granted, checked at each step, and written to a log that cannot be edited after the fact. You choose whether your data lives in the EU or the US, and HQ is built under GDPR and the EU AI Act from the first line. → Permissions are scopes; the record is the Audit API.HQ is invite-only while we onboard teams. Slack and the browser extension are live today, with Microsoft Teams on the way.