Quickstart
Waken gives your agent its own machine in the cloud. It keeps its files, its installed packages and its conversation. It sleeps when nobody talks to it, and wakes in a fraction of a second on the next message. This page takes you from nothing to a running agent in about two minutes.
1. Install the CLI
curl -fsSL https://waken.sh/install.sh | sh
The script downloads one binary for macOS or Linux (and WSL on Windows), checks its checksum, and puts it in ~/.local/bin. Prefer to stay in the browser? Everything below also works from the dashboard.
2. Sign in
In the dashboard, open Developers and create an API key. Then:
waken login# Paste your API key: wk_...# Signed in. 0 agent(s) on this account.
3. Deploy a folder
An agent is a folder: instructions, files, scripts, whatever it needs. Waken runs a harness in it (Claude Code by default) with your own model key.
mkdir bookkeeper && cd bookkeeperecho "You keep my invoices in order. Files live in ./invoices." > CLAUDE.mdexport ANTHROPIC_API_KEY=sk-ant-...waken deploy# Deploying bookkeeper...# Agent 4ec0b05a running in 0.12s# 12 files uploaded
Your key goes to the agent's memory, never to its files, and it is never shown again. No Anthropic key? Use another provider: see Models and keys.
4. Talk to it
waken chat "What can you do for me?"# or interactivelywaken chat
Close your laptop. The agent stays where it is. After five minutes without activity it goes to sleep, and the next message wakes the same machine, same files, same conversation, in well under a second.
From your phone too: connect Telegram, Discord or Slack in the dashboard, under Integrations, and every message you send there goes to your agent. Same agent, same memory as the terminal and the dashboard. See Channels.
5. Rewind when it breaks something
A checkpoint is taken every 15 minutes while the agent runs, and you can take one yourself before anything risky. A rewind restores the whole machine: files, packages and the conversation. It cannot undo what already happened outside the machine, like an email sent or a payment made.
waken checkpointwaken checkpointswaken restore # the latest onewaken restore 20260920-161728-38e2
Ready to deploy agents
The Templates page of the dashboard lists agents with a name and a harness already chosen. Deploy one, then send it its role as the first message. Here are the roles, so you can adapt them:
| Agent | Harness | First message |
|---|---|---|
| inbox-triage | claude-code | You are an inbox triage agent. Every time I send you files or a mailbox export, sort them into folders by topic, keep a running index in INDEX.md, and reply with a five line summary of what arrived. |
| bookkeeper | claude-code | You are a bookkeeper. When I send invoices or receipts, file them under YYYY/MM, add a line to ledger.csv (date, vendor, amount, category), and tell me what looks missing or duplicated. |
| market-watch | hermes | You are a market watch agent. I will give you a list of sources in sources.md. Each time you run, read them, compare with notes/yesterday.md, write notes/today.md, and reply with what changed in three bullet points. |
| support-notes | codex | You are a support assistant. When I paste a customer conversation, extract the question and the answer into kb/, then draft a reply I can send. Never send anything yourself. |
| repo-gardener | cursor | You are a repository gardener for the code in this folder. Each time you run, pick one small improvement (an outdated dependency, dead code, a flaky test), make the change on a branch, and describe it in one paragraph. |
| night-qa | openclaw | You are a night QA agent. When you run, execute the test suite in this folder, investigate any failure, and write report.md with what failed, the likely cause and a suggested fix. |
A Deploy button in your README
Anyone who clicks it lands on the creation form with the name and harness filled in, after signing in. Replace the name and the harness with yours:
[](https://waken.sh/deploy?name=my-agent&harness=claude-code)
Harness values: claude-code, codex, openclaw, cursor, hermes, zeroclaw. The name must be lowercase letters, digits and dashes.
Where to go next
- CLI: every command.
- HTTP API: one agent per customer, from your own backend.
- Desktops and MCP: give an agent a real screen and a browser.
- Container images: run your own agent instead of a harness.