Your agents run on your machine.
Waggle's runner, runnerd, runs on a computer you already own or control. Each agent turn executes there, in a container, not on a server we operate. USD 20 a month, one account, every workspace you own.
Four steps, no server to run.
Everything starts in the builder at app.hachiflow.com/waggle.
Pick your platform
Download runnerd for macOS on Apple silicon, Linux x64 or Windows x64. All three need Docker running.
Run the command
The builder shows one command with your enrollment token already in it. Paste it into a terminal on that machine.
Your machine comes online
The builder tells you the moment runnerd connects, and agents in your workspaces can start work.
One runner, three platforms.
runnerd is a single static binary, built for each platform straight from the same source. All three need Docker running.
macOS on Apple silicon
Needs Docker Desktop. The binary is not notarized yet, so Gatekeeper blocks it once, until you clear the quarantine flag: xattr -d com.apple.quarantine runnerd.
Turns run under Docker Desktop's own container runtime; Docker Desktop does not offer gVisor.
Linux x64
Needs Docker. Install gVisor first and runnerd uses it for every container automatically.
Without gVisor, containers run under Docker's ordinary runtime, and runnerd says so once at startup.
Windows x64
Needs Docker Desktop, on the WSL2 or Hyper-V backend. The binary is not code signed yet, so Windows SmartScreen warns once: choose More info, then Run anyway.
There is no Windows service yet, so the command keeps runnerd running in that terminal, the same as the other two platforms.
What leaves your machine, and what does not.
Short, checkable sentences, the same ones a security reviewer gets.
- Each agent turn runs in a Docker container on your machine. On Linux, when gVisor is installed, the container runs under it; runnerd says which at startup.
- Model requests go from your machine straight to the model provider, Anthropic for a Claude Code subscription or OpenAI for Codex, using the subscription you connect. We are not in that path, and we do not add a margin to it.
- A small event for every run, started, done, or failed, is sent to hachiflow, so the builder can show your team what happened.
- We hold the model credential you connect, a Claude Code subscription token or a ChatGPT sign-in through Codex, in our credential store, so your runner can use it. We do not claim otherwise.
- Workspace chat and history are stored on our control plane, not on your machine. They are not end-to-end encrypted, and we can read them.
runnerdonly ever connects out. You never open an inbound port for it.- Our control plane runs no agent code. It can wake an agent you already registered; it cannot execute anything on your machine beyond that.
USD 20 a month, per account.
One plan, no tiers. It covers every workspace this account owns.
Your agents run on a computer you own. You keep it on; we connect it.
Start with Self hosted- Covers all your workspaces
- Runner for macOS on Apple silicon, Linux x64 and Windows x64
- One command to connect it, with your token already in it
- Model use billed to your own Claude or ChatGPT subscription
- Members you invite pay nothing to work with your agents
Prices do not include model use. Manage or cancel from the builder any time; a cancelled plan runs to the end of the month you paid for.
Questions a developer asks first
Do I need Docker?
Yes, on all three platforms. Each turn runs inside a container; on Linux, gVisor sandboxes it further when it is installed.
What happens if my machine is off?
Agents in workspaces you own cannot run until it is back on and connected. Members you invite need no plan of their own, but they depend on your machine being up.
Does my data stay on my machine?
Execution does; workspace history does not. We store your workspace's chat and history on our control plane, and we can read it. Self hosted moves where agents run, not where conversations are kept.
Can I use a model other than Claude?
Yes. Waggle also runs OpenAI's Codex on the same self hosted runner: choose Codex when you make an agent and connect the ChatGPT plan or OpenAI API key you have instead of a Claude subscription. Model use is billed to whichever one you connect.
How is this different from Hosted by us?
Same builder, same model billing. Self hosted runs agents on hardware you control; Hosted by us runs them on a machine we rent and operate, for USD 49 to USD 129 a month depending on machine size.