U2OS — the operating system for your digital self: a vault of Markdown files you own, plus an agent that acts…
FAQ (answers come from our records)
How do I install it?
We did not find an install command for this repository (our install coverage is only about 4%). Check its README or releases page.
Who is it for?
We filed it under these tags: 开源项目. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
repository chrisrobison/u2os;first catalogued 2026-10-15;last seen 2026-10-16。There is no official certification or quality rating here — we record sources and dates only.
U2OS
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
U2OS is **the operating system for your digital self**. It consolidates who you are into **files you own**, and **acts on your behalf** within the authority you delegate.
> Observe → remember → anticipate → act → observe outcome → learn.
It is not a chatbot. The digital self is a **vault of plain Markdown files** that you can read, edit, version and carry anywhere:
- who you are (me.md)
- the people and projects in your life, and what you have promised
- the standing routines you want carried out, and skills describing how you want things done
U2OS indexes that vault, watches the accounts you connect, and runs your routines unattended. Every action goes through a policy engine that sits outside the model. The agent is just a tool that uses this information, and the model is replaceable infrastructure. The name means "the second you" plus "operating system".
Start with the overview, then the vault and routines. ADR 0007 records why the vault is the centre, and PLAN.md is the roadmap. ADR 0010 sets the direction for a lean core with bundled add-ons, and add-ons specifies the contract. PROMPT.md is the historical product specification, not onboarding documentation.
How it works
- **The vault is the self.** Files are the authority for what they say. SQLite holds an index of them plus runtime state (sessions, queue, runs, audit, connector caches), and vault-sourced memory can be rebuilt from the files.
- **Routines act for you.** A routine file names a trigger (a daily time, an interval, or an event such as email.received) and an instruction in your own words. It runs without a chat, through the same planner → policy → queue → audit path as everything else, and never with more authority than your policy delegates.
- **Policy is outside the model.** Reads and drafts can be automatic. Sending, rescheduling and other consequential actions ask you or are blocked, according to policy. Model output, routines, feedback and voice confidence cannot loosen that.
- **Privacy is separate from permission.** Each fact carries a classification (public, personal, private, sensitive), and the data-processing policy decides what may reach a local or remote model.
- **Everything is explainable.** Actions keep their reasoning summary, policy rule, model identity and the memory that informed them.
The browser is a client of a persistent Node.js service. Closing it does not stop sync, routines, triggers or the agent.
Project status
U2OS is a working **pre-alpha**, not a production product. Everything below is implemented and tested with fixtures. **It has not yet been validated in daily use with real accounts and a real model**; that is the next milestone (PLAN.md, Milestone B).
**The owned digital self**
- Markdown vault with YAML frontmatter: me.md, people/, projects/, commitments/, routines/. It is indexed on start and on every change, and its location is set with U2OS_VAULT or vaultDir.
- Vault facts are explicit memory with vault: provenance. Edits supersede, removals soft-delete, and other sources are never modified.
- Per-file classification and sensitive_keys control what can reach a model.
- npm run vault:export moves existing database memory into vault files, bound to the same records with id:. Memory-UI edits are written back to the files, policies.yaml in the vault states what U2OS may do without asking (an invalid file fails closed), and journal/ records what it did on your behalf.
- Standing routines run on daily, interval or event triggers. Each slot runs once, even across restarts, and runs go through policy and approval, with a runaway limit. Routines can use vault **skills** (Markdown instructions) such as the example Job Hunter.
- **Coding agents**: hand a software task to the Codex CLI or Claude Code you already use. U2OS launches the official tool and never holds its credentials (coding agents).
- Tools from **MCP servers** declared in your vault's mcp.yaml run out of process, behind the same policy gate (MCP tools). The first-party **job-hunt server** searches Greenhouse and Lever boards and applies in a headless browser, with an application ledger in your vault (job hunting).