# ALINA OS: an AI-native operating system > A computer with a resident intelligence. ALINA OS boots into a conversation. A fleet of agents works across the whole machine. It remembers with rules, acts where you can see it, submits to review by a model from another lab, and runs on a NixOS system declared entirely in text. Built by [ControlStackAI](https://controlstackai.com). Source is private while the public distribution is separated from personal configuration. Film (4:31): https://youtu.be/VC16TgyUrMc ## Where it's going ALINA is becoming an always-on intelligence personal to one human, with memory, authority and identity held by an OS the person owns. It grows three memories: episodic (Loom, live), semantic (shared memory curated nightly by a council of three labs, live) and parametric (Engram, a personal model with memory in its weights, in research). The personal model calls frontier models when it needs them, earns each role on the person's own work, and will run at home on an ALINA box. Full vision and the M0-M11 road: https://alinaos.com/vision/ ## New in October 2026 ### Loom: any harness picks up any session Status: P0 ingest live; P1 presence built and tested; P2 takeover next. Details: https://alinaos.com/loom/ Claude Code and Claude Desktop, Codex CLI and Codex Desktop, pi, Antigravity and OpenClaw all feed one ledger that the operating system owns. A thread becomes an OS object: an append-only ledger of what happened, a worktree, and a lease that records who is driving. The agent apps become lenses on it. - Full mesh, no hub: one shuttle per harness translates to a canonical schema, so every pair works without a pairwise translator. Loom never runs inside a harness. - Effects, not tool calls: the ledger records "this file changed by this diff" and "this command exited 1". - Trust travels with the text: web pages, tool output and relayed chat are tagged untrusted and quoted as data. - MCP tools in every harness: `loom_list`, `loom_read`, `loom_search`, `loom_watch`, `loom_post`. - OpenClaw sessions are ingested but OpenClaw is never given the read-everything lens. ### Shareable NixOS: a teammate's laptop from one command Status: starter released in openclaw-nix 2026.10.3-r4; team layer in progress. Details: https://alinaos.com/nixos/ The configuration splits into product, team, machine and personal layers. `nix run github:alinaos-org/openclaw-nix#init -- ~/laptop` generates a person's own flake with their identity, encrypted secrets and their own sign-ins. Configs hold only `${VAR}` references; a reference without a key fails the build and names the key. Sign-ins are never shared. CI generates a laptop and boots it in a VM on every change. `openclaw-doctor` checks environment, gateway, secrets and models after the switch. ## How it works: four questions ### It knows: memory with rules - Persistent, partitioned memory shared by every agent and coding harness. Recall is scoped by audience and refuses requests whose context can't be proven. - A nightly Memory Council: models from three AI labs propose and cross-examine candidates; a rotating judge decides. - No erasing: the curator cannot delete records, including its own. ### It acts: the desktop is an API - Agents launch, focus, move and notify on the Hyprland desktop through a broker. - Every window signed: each agent works in its own workspace; every window carries its name and border color. - Every desktop call is logged; when the session locks, the fleet drops to read-only. - Realtime voice interrupt without losing agent capability. ### It's checked: review before action, proof after - Cross-model review first; verdict plus blast radius decide whether work lands reversibly or waits for a person. - The desktop stages the review in the right app on the right monitor with an approve or request-changes control. - Consequential actions need exact-action approval, run once, and leave a signed receipt. Replays are rejected. ### What it stands on: the whole machine is a diff Disk layout, desktop, services, agent configuration and policy are declared in a NixOS flake. Agent changes to the OS are reviewable diffs. Rebuilds are atomic; rollback is one generation away. ## The agent fleet Specialist agents on an OpenClaw control plane, each with its own workspace and durable memory: ALINA (resident assistant), ATLAS (strategy), ORACLE (whole-corpus reader), MODELER (MATLAB/Simulink), MUSE (creative), HERALD (publishing), STUDIO (review-gated video), SPECTER (offensive security), DISSECT (reverse engineering), NEXUS (networks), LEDGER (finance), LIFELINE (emergency recovery). Channels: desktop voice, web Control UI, iPhone app, Slack, Discord, self-hosted Buzz relay over Nostr. Models: Claude, GPT, Gemini, Kimi, Qwen, routed per agent and workflow. Workflow profiles (plan-execute-verify, review council, chain-of-verification) are chosen per session by the person. ## Dark Factory: autonomous work that fails closed Issue ready → worker (one lab, isolated workspace) → reviewer (a different lab) → proof packet (commit, tests, token accounting; deployed must match source) → done. Missing proof blocks the run. Phase 0 passed on 2026-10-14: a real issue went from Ready to Done autonomously. ## Built like a system - 290× less CPU after the status view moved from a polling script to an event-driven native daemon. - 5 agent harnesses in one OS-owned session ledger. - 3 AI labs on the nightly Memory Council. - 1 compile per OpenClaw release, built in CI and cached. ## Public repositories - https://github.com/alinaos-org/openclaw-nix - https://github.com/alinaos-org/engram - https://github.com/alinaos-org/linux-voice-cockpit - https://github.com/alinaos-org/buzz-openclaw - https://github.com/alinaos-org/agent-installer - https://github.com/alinaos-org/alinaos-showcase