FAQ

AIOS (Personal AI OS) is a desktop application — Tauri v2 shell (Rust + React + TypeScript) wrapping a local FastAPI/Python sidecar — that runs a research, knowledge, and automation workstation entirely on your own machine. It has 21 tabs covering chat, research, document/image ingestion, a coding-agent bridge, a data sandbox, and more, all backed by a local SQLite database and a local ChromaDB vector store.

No. AIOS itself only runs as a local desktop app on your machine — there is no hosted version. The only cloud-side component is aios.graphenlabs.com, a separate Next.js marketing and account/license portal used for checkout and license management; it never processes your documents, chats, or research data.

Free is a permanent tier, not a time-limited trial, but it's capped: a 500-document knowledge base limit, 3 agent dispatches per day, and a fixed (non-configurable) cost ceiling on model spend. Core chat, research, the knowledge base, the Discovery graph, and MCP-server mode are all included at no cost.

Not strictly. AIOS runs local models through Ollama with a single-worker scheduler if you have GPU-capable hardware, but you can also bring your own API key (BYOK) for a cloud provider like Claude or OpenAI, or use the Claude Code fallback path — either way, a local GPU is not a hard requirement.

AIOS's onboarding flow detects whether Ollama is reachable and, if it isn't, guides you toward installing it or using BYOK / the Claude Code fallback instead. Local inference features simply won't work until either Ollama is running or a provider key is configured.

Windows is the confirmed, tested path today — AIOS ships as a signed MSI and an NSIS installer, both Windows installer formats, built and verified through a Windows-first packaging track. macOS and Linux are not currently packaged or verified; the project's own packaging docs explicitly scope macOS work as not-yet-started rather than shipped, so don't treat cross-platform support as available yet.

Free covers full local chat, research, and knowledge base with usage caps (500 documents, 3 agent dispatches/day) and a fixed cost ceiling — MCP-server mode is included on Free too. Individual is a paid subscription that removes those caps — unlimited knowledge base and agent dispatch, multi-agent support, and a full personal spend ledger. Enterprise is per-seat/month and adds shared workspaces, SSO/SAML/SCIM, audit logs, RBAC, and an admin console.

Yes — Free is designed as a permanent tier, not a trial, and no core capability is withheld to force an upgrade; the tiers differ on usage ceilings (document caps, agent dispatch limits) and organizational features (SSO, shared workspaces), not on whether basic chat, research, or the knowledge base work at all.

Billing runs through Dodo Payments, not Stripe — Stripe doesn't onboard India-incorporated businesses as merchants of record, so Dodo handles checkout and payment processing for the Individual tier. Authentication (sign-in/sign-up) is handled separately by Clerk and is not coupled to billing.

Refund policy specifics aren't published here yet — check back on our Terms page once it's live for the current policy.

By default, no — documents are embedded and stored locally in SQLite and ChromaDB, and local inference through Ollama never leaves your machine. Data only leaves if you explicitly trigger it: using a BYOK cloud provider key, running a Web tab search, or connecting an external MCP client to AIOS's MCP server.

AIOS can act as an MCP server, exposing its research, memory, and web tools to external MCP clients like Claude Desktop, Cursor, or Codex CLI — this is free on every tier, including Free, and works even without a local GPU or a paid AIOS license. It's private in the sense that AIOS itself stays local and only responds to the MCP client you've explicitly connected; it does not phone home or share data with Graphenlabs as part of that connection.

License tokens are signed with Ed25519 by the account portal at purchase time and verified entirely offline by the desktop app against a public key baked into the build — there's no runtime check-in to a license server required for AIOS to know what tier you're on.

Across its 21 tabs: chat with local or cloud models, run semantic research with cited Q&A, do live web search, chat with your own uploaded documents, describe/OCR/ask about images, recursively index local folders, automate a headless browser, dispatch a real coding agent against a repo, run a sandboxed data-analysis script, maintain a knowledge base and context library, generate weekly plans, spec and draft content, and explore a Discovery graph of related findings — plus model management, prompt/persona libraries, and hardware/system telemetry.

The Orchestrator tab dispatches a real external coding agent (Claude Code or Codex) against a local repository: it generates a plan first, requires your explicit approval before running anything, and only operates inside an allow-listed set of repo paths with a scrubbed environment. It's not a sandbox in the kernel sense — the agent has real filesystem access within the allow-list — so the allow-list itself is the control that matters.

Yes — BYOK (bring your own key) for cloud providers including Claude and OpenAI is supported on every tier, alongside local Ollama models. This is also the path that lets you use AIOS meaningfully without a GPU.