Agents on your own hardware
Coding agents you can let loose on real code.
Every agent runs in its own VM on your own machines. It can't see your host, your model keys or another team's work. Budgets stop runaway spend, and every action lands in a tamper-evident log.
How it works
A task in, a review out.
Every card on the board flows through its lanes. An agent picks it up in a fresh VM, does the work, and another one reviews it before it reaches you.
# install on a Linux machine with KVM $ ./install.sh --profiles "claude codex" # more machines join as nodes $ ./install.sh --nodes 7790
Why fluwie
Agents that can't step out of line.
Isolation
Each agent gets its own virtual machine, not a shared container. Model keys stay outside, behind a proxy; your host, your LAN and other agents are out of reach.
Control
Hard budget stops per task, project and install. A person signs off where your workflow says so, and every action is in a hash-chained audit trail.
Ownership
Your machines, your Git forge, your models, your registry. fluwie runs no service you depend on: nothing leaves unless you send it.
What an agent can't do
Safe by construction, not by promise.
An agent in fluwie works inside limits you set. These are not settings it can talk its way around.
- Reach your model keys. A proxy outside the VM adds them to each model call; the agent never sees them.
- Reach your network. It can't see your host, your LAN, your VPN or other agents' VMs, unless you list an exception.
- Spend without limit. Hard budget stops per task, project and install, checked before every model call.
- Act unseen. Every action is in a hash-chained audit trail you can verify.
- Merge on its own. If your workflow has a gate, a person approves before anything lands.
Features
Everything a team of agents needs.
Board & workflows
A kanban board whose lanes are steps: write a spec, plan, build, review, test. You decide where a person signs off.
Roles & skills
Give agents a role (developer, reviewer, tester) and the skills it needs. Reuse them across projects.
Security review
A workflow step where an agent checks each change for security issues before a person sees it.
Fast VMs
A VM starts from a ready snapshot in about a tenth of a second, at about 10 MB each. Ten agents start as easily as one.
Nodes
More machines join as nodes and run agents for the first one. Scale out when the board fills up.
A console everywhere
Board, terminals, documents and notifications in the browser, and on your phone as an app.
FAQ
Questions
What is an agent in fluwie?
A coding agent running in its own microVM, with a repository, a role and its skills. It works on a task from the board, or on what you tell it in its terminal.
Where does it run?
On your own x86_64 Linux machines with KVM. More machines can join as nodes, and the control plane can also run in Kubernetes.
Which agents and models can I use?
Claude Code, Codex, Cursor, OpenCode, pi and Hermes are built in. Point them at hosted models or your own model servers.
Is my code safe?
Every agent is in its own VM on a private network with no ports open to it, and it can't reach your LAN, your VPN or other agents unless you allow it. Model keys never enter the VM, secrets live in a keyring, and every action is in an audit trail you can verify.
Can our security team sign this off?
That is what it is built for: it runs on your own hardware, keeps your code there, and needs no outside service. We'll walk your security team through the isolation, network and audit model.
How do I get it?
fluwie is in early access. Send us a note and we'll get you set up.
Put your agents to work.
fluwie is in early access. Tell us what your agents should work on, and where it has to run.