Twenty terminal windowsSSH into every boxtmux and hopeLosing the transcript
Speed is work that overlaps
A coding agent spends most of its life waiting — on a build, on a test run, on a model. Watch one and you wait with it. Run ten and the waiting overlaps.
up to10
Claude sessions at once
up to20
sessions on Claude Max
Multiple
Claude accounts supported
Waiting on each other0%
10% parallel work
Running together0%
Done
90% parallel work
The same fifteen tasks, five machines, one clock. The right-hand schedule finishes and the clock stops; the left is left standing at a third of the work — which is what the red figure is.
Flow Coding
This is the new flow coding: voice and parallel Claude. You talk, several machines work at once, and you read results instead of watching a spinner. Nothing here makes one agent faster — a build takes what a build takes. What changes is how much of that time is yours.
The schedule on the right needs machines to run on. It does not need new ones — a desktop, a spare VM and a container on a box you already own are three more machines, and each takes about two minutes.
A PC you already use
Linux, or Windows with WSL2
Your own desktop is a fleet machine. The agent host is a single executable that carries its own runtime, so the machine needs the claude CLI and nothing else — no Node, no npm, no dependencies to keep current.
Register the machine in the console and press Remote
Carry the file it gives you to the PC and run it
The release is linux-x64. On Windows that means inside WSL2, and each WSL distribution is one machine rather than the PC being one.
Not tested by us — the pieces are the same, the route is not one we run.
Containers on one box
Docker, and a volume per machine
The densest option: several machines on one host, each in its own container with its own workspace and its own credential. This is how our own development boxes run — one image, one container per machine, all dialling the same console.
Install into the image, or into a running container
Mount a volume at the state directory before first start
The state directory MUST be a mounted volume. The claim is spent on first start and the token is written there — a state directory that does not survive being recreated is a machine that cannot come back, and a container is recreated often.
VMs on existing hardware
Any Linux VM, 2 vCPU upwards
The ordinary case, and what most of our own fleet is. Spare capacity on a hypervisor you already run becomes as many machines as you have cores to give them; each one is independent, so one machine’s build cannot slow another’s.
Register them all at once — the console issues one claim each
Run the one script it writes, from a box that can reach them
Machines dial out to the console over one outbound WebSocket. Nothing listens on the VM, so there is no inbound port to open and no firewall rule per box — but the VM does need to reach the console.
Every machine holds its own credential, created on the machine and never sent to the console — which is why adding one is cheap and losing one costs nothing but the machine.
Talk to it
Click the microphone and speak. Words appear in the compose box as you talk, transcribed at each pause and appended — never revised underneath you.
It knows your machines
Machine names, repository names and the session's own title are sent as bias, so vm-03 and your repo come back spelled properly instead of phonetically.
It gets out of the way
Click into the box and the microphone pauses. Two seconds after you stop typing it resumes, appending to what you typed rather than overwriting it.
It never sends on its own
You read what is in the box and press Send yourself. A spoken word can never approve a destructive call — dictation puts text in a box, it does not answer approvals.
Multiple voice models supported
Three transcription providers, chosen per organisation on the Voice page. They differ by more than price — how each one is told about your machine names changes what comes back, so the console lets you record once and compare them side by side before committing.
GroqWhisper large v3 turbo
~US$0.04 / hour of audio
The cheapest by a distance, and the fastest. Custom words go in the prompt.
OpenAIGPT-4o mini transcribe
~US$0.18 / hour of audio
Four times Groq's rate; worth having where the key already exists for something else.
GoogleChirp 3
~US$0.96 / hour of audio
The only one that takes custom words as data — a weighted phrase set rather than a sentence to bear in mind.
Billed by the second, by you, on your own key — the console holds the key and never resells the capacity. Every provider is checked from the settings page before it is trusted with a recording.