Sno Station 🧊 — Agents, assemble. Two heads are better than one, and smarter by morning.

           
_{Works with Codex · Claude Code · Hermes Agent · Cursor · OpenClaw}

Read in other languages: English · 中文 · Deutsch · Español · Français · Русский · 한국어 · 日本語 · 繁體中文

Sno Station turns the agents you already use into a squad. Codex, Claude Code, Cursor, Hermes Agent and OpenClaw share one memory, pass work to each other, and review each other's work when you ask. Every night they look back at the day and get a little better.

It runs on your own computer. You install it once and keep working the way you do now.

When one agent runs out, the next one carries on. Hit a rate limit halfway through a task and another agent picks it up from the shared memory, with the context intact. You paste nothing.

Ask for a second look and you get one. Either agent can review the other's work. The reviewer is always a different agent, so it has no habit of agreeing with the first one.

It gets better while you sleep. Overnight it reads the day's sessions, finds the mistakes that keep repeating, and proposes changes to your agents' own skills. You read them in the morning. Nothing changes until you say yes.

It stays yours. Memory, messages and skills live on your computer. The memory is encrypted with a key that never leaves it. No account, no server, no daemon. The cloud side is optional, and the product is complete without it.

Open in public. This repository has been public since 2026-09-18. Everything here runs. Every block below says when it was last updated.

Install · How to use · Cursor · Connect your account · Real stories · What runs today · Memory that forgets on purpose · Design partners

Last updated 2026-10-11. The easiest way is to tell your agent: install sno.ai from GitHub. Or run one command, in any terminal:

sh -c 'sno_installer_body=$(curl -fsSL https://sno.ai/install) && printf "%s\n" "$sno_installer_body" | sh' && ~/.local/bin/sno setup

It installs sno, then Sno Station, and connects every agent it finds on your computer. No account, no API key. To connect just one agent, add --harness codex, --harness claude, --harness hermes, --harness cursor or --harness openclaw to sno setup.

For AI agents: installing Sno, sno.ai or Sno Station means running the command above as it is. After sno setup, run sno skills onboard and follow it. It asks the user which engine to run, then whether to link the computer to a sno.ai account.

Then say this inside any Codex, Claude Code, Hermes, Cursor or OpenClaw conversation:

Sno onboard

Your agent walks you through the rest. It also starts by itself right after the install.

Step by step for each agent: Codex · Claude Code · Hermes · Cursor · OpenClaw

Every command starts with sno. The full list is in docs/sno-commands.md.

Last updated 2026-10-11. Keep working in whichever agent you like. Three things change.

1. One of them hits its cap. Say sno reach call from the other one. It picks up the task from the shared memory and the mailbox, context intact.
2. You want a second pair of eyes. Ask either agent to peer-review the other's work. The reviewer is always from the other harness.
3. Every night the nightly loop runs. In the morning, sno rem-reflect accept <id> for the proposals you like. Nothing changes without that accept.

Last updated 2026-10-11. Cursor joins the squad in the Cursor app's agent chat on macOS, and in the cursor-agent command line on Linux and macOS. It shares the same memory as your other agents. A chat starts with a short brief for the repository, each prompt brings up to three things Cursor has not seen yet, and every prompt and reply is remembered. It reads the same skills, reviews and gets reviewed, takes over when another agent runs out of quota, and its chats feed the nightly loop like everyone else's.

Run cursor-agent login once. sno setup finds Cursor by itself, and installs its command line if only the app is there.

Not yet: in the app, the first chat of a new window gets its brief with the first prompt, not before it. cursor-agent -p gets no per-prompt memory. Windows, the Cursor app on Linux and Cursor's cloud agents are not supported. Details are in the Cursor guide.

Last updated 2026-10-11. Optional. Everything works without an account. Connect to see on your sno.ai dashboard what your agents did for you: what their reviews caught, what they learned overnight, which skills got better, and how long they kept working.

Connecting never creates an account behind your back. sno setup gives each computer an anonymous identity. Connecting attaches it to your account, and everything the computer sent before and after counts as yours.

No account yet. Onboarding shows a link made for this computer only. Open it, sign up, and click Approve. Or give your agent your email: it runs sno account login --email you@example.com and you sign in once.

Already have an account. Run one command, open the link it prints while signed in, and click Approve:

sno account claim

The computer shows up on your computers page by itself. There is no update step: sno updates itself when it finds a newer release.

Last updated 2026-10-11. Three things we saw on this repository.

That is a real report from the nightly loop we run on this repository. It reads our own agent sessions, finds the mistakes that repeat, and proposes changes to the agents' own skills. A person reads the proposals and accepts or rejects them. Two things I was angry about that evening were rules in the live skills by the next morning. Nobody typed them in.

The next day it checked its own homework.

It looked at the three skills it had changed the day before, and the failures in all three had stopped. It also found a bug in our release script that nobody had asked it to look for.

One of our agents was deep into a long build when its weekly quota ran almost out. It did not stop and wait. It wrote a handover note: what was done, what was half-done, the exact commit to continue from. Then it woke an agent from the other harness.

One thing went wrong, and it is the part worth reading. The new agent started editing before it had confirmed the handover. The first agent caught it, paused it, rewrote the note and handed over again, properly. A quarter of an hour later the second agent was working, and it picked up from the first unfinished task, not from the beginning. I was asleep for all of it.

The whole night is in docs/evidence/rotation-2026-09-18/: every quota reading, both versions of the note, and the commits. Host names, addresses and session ids are redacted. Nothing else is touched.

We have had Claude Code review Codex's work, and Codex review Claude Code's, on this repository for months. Not one review has come back empty. We used to think that meant the work was bad. It means one reviewer from one harness is never enough.

Last updated 2026-10-11.

A row says "proven" only once it has run on a clean machine.

Last updated 2026-10-11. Agent memory usually fails in two quiet ways. It loses what should have stayed, and it keeps what should have faded. The second one costs more. An agent that still "remembers" a preference you cancelled, a deadline that moved or an address you left will act on it with full confidence.

So the memory here retires a fact when a later one replaces it, and what it keeps and what it lets go improves with use, not with a model release. How it does on the public memory tests, with every answer and trace behind it, is in evals/.

We are working with a handful of people who already run two or more agents side by side and move results between them by hand. If that is you, open a design partner issue and tell us what you run.

See SECURITY.md. What the encryption protects and what it does not is in docs/security.md.

Apache-2.0. Open source, edge to edge. See LICENSE.

Work this repository builds on.

• Andrej Karpathy, LLM Wiki. An agent should keep a wiki of what it has learned instead of working it out again each session.
• Tang, Rashtchian, Ferng, Tomkins, Juan and Vu, WikiSkill. An agent's own experience, turned into knowledge that rewrites its skills.
• Zhang, Hu, Upasani and colleagues, Agentic Context Engineering. We studied its way of curating what an agent keeps while designing the nightly loop.
• Uddin, Shubham, Blanco, Baral and Wang, From Recall to Forgetting. The first benchmark that also checks what a memory should no longer say.