Now in early access — 7-day free trial, no card required →

Your AI workforce.Your infrastructure, or ours.One control center.

Your Claude, Codex or Copilot work your tickets through your own delivery flow: plan, review, build, verify, your gate, merge. Your architects draw that flow and we sit with them to do it. Every run gets a disposable DevBox on your fleet or our cloud, and when one is unsure it stops and asks.

Pilot it on three tickets See how it works →

Your Claude, Codex or Copilot, not oursWritten with your architects, supported by AptiveAIOpen coreNothing merges past your gate

Every engineer runs AI their own way, and nobody can say what it did last night. See what breaks →

One ticket, end to end

Watch one ticket go all the way around.

You assign #482 from the board you already use. It gets its own throwaway machine, writes the code and runs the tests. Then it stops at your gate as a draft pull request and waits. The right half of the ring is the agent’s. The left half is yours.

Outside the ring: today. Each engineer's own tool, own habits, own machine. Nothing shared, nothing kept.
On the ring: every ticket, every night. The board assigns, a disposable DevBox builds, results are verified before anyone reads them.
One stop is yours. Nothing merges until it passes the gate you defined. Approve, send back, or take the machine.

one engineer sits with itstops when the laptop closeson a real laptop, unsandboxedno record when it is done

Assigned tonight · Waiting at your gate · Merged when you say so

maya@mbp — zsh — 92×24

maya@mbp ~/core $ claude -p "fix #482" --dangerously-skip-permissions▸ reading src/api/invite.ts▸ editing 3 filesmaya@mbp ~/core $ npm test✓ 212 passed (14.2s)maya@mbp ~/core $ git commit -am "wip"[main 4f1c2a9] wip

dev@thinkpad — bash

dev@thinkpad:~/work/core$ codex exec "migrate notification kinds to the registry"▸ thinking…▸ applied patch (+41 −12)dev@thinkpad:~/work/core$ npm run lint✗ 3 errorsdev@thinkpad:~/work/core$ claude "why is this failing"▸ reading logs…

sam@studio — fish

sam@studio core/ ❯ claude --resume▸ session not foundsam@studio core/ ❯ claude -p "…again, from the top"▸ re-reading context (2m 10s)▸ editing client.tssam@studio core/ ❯ git pushTo github.com:acme/core.git

ana@air — zsh

ana@air ~/core $ copilot suggest "tests for #491"▸ wrote 4 specsana@air ~/core $ npm test -- invite✓ 4 passedana@air ~/core $ git pushEverything up-to-date

lee@desk — zsh

lee@desk ~/core $ codex exec "review the diff"▸ 2 commentslee@desk ~/core $ claude "address the comments"▸ …▸ editing 2 fileslee@desk ~/core $ npm test✓ 214 passed (15.0s)

kim@mini — bash

kim@mini:~/core$ claude -p "add retry hint to the client"▸ editing client.tskim@mini:~/core$ npm run lint✓ cleankim@mini:~/core$ git commit -m "retry hint"[main 9b7e01d] retry hint

The Status Quo

Every developer on your team has their own AI workflow.

Twelve engineers, twelve private ways of using Claude and Codex. Nobody sees what ran overnight, what it touched, or what it cost. You did not choose that. You were handed an engine.

With ClusterCode

One workflow the whole team runs.

Same tickets, same agents, one board. Every run visible from the moment it is assigned.

#482 · claudeMaya's laptop

#488 · codexDev's terminal

#491 · claudeWorking · 14m

#493 · copilotTests running

#470 · claudeMerged 02:14

#495 · codexReady for dev

#489 · claudeWorking · 41m

#486 · claudeReview pass

#479 · codexAwaiting you

#468 · copilotMerged 01:52

#497 · claudeReady for dev

#490 · codexWorking · 8m

#484 · claudeBrowser check

#481 · copilotAwaiting you

#465 · claudeMerged 00:31

Levels of autonomy

Your company is arguing about this scale.
In ClusterCode, your place on it is a gate count.

Researchers at Columbia describe five levels of agent autonomy, and they define each one by the role you play rather than by how clever the agent is. That is the useful part: your level is a decision you make, not a capability you buy. In ClusterCode that decision is where your architects put the approval gates. Take one out when the evidence says you can. Put it back the day it says you cannot.

You are probably here
LEVEL 1
Operator
You direct the planning. The agent supports on demand.
An engineer sits with the agent in their own terminal. The work stops when the laptop closes.
a person at every step
LEVEL 2
Collaborator
You and the agent plan and execute together.
The run is on the board, and you can take the live terminal mid-run and hand it back.
3 gates
LEVEL 3
Consultant
The agent leads. You give feedback and guidance.
It writes the plan and waits. Approve it, or send it back with a note, up to three times.
2 gates
Where most teams settle
LEVEL 4
Approver
You approve consequential actions. Otherwise minimal involvement.
Nothing merges until you say so. One stop, at the pull request, with the diff and the cost.
1 gate
LEVEL 5
Observer
You monitor only, with no intervention except an emergency stop.
For a kind of work you have watched succeed a hundred times. It still stops when it is unsure.
0 gates · asks on doubt

Each strip is one run: plan · build · verify · merge. A filled dot is a human gate.

You do not pick one level for the company. Each kind of work has its own workflow, so each carries its own: dependency bumps can sit at four while anything touching customer data stays at one. Your architects place the gates, we sit with them to do it, and thirteen templates mean nobody starts from a blank page.

The five levels, and how to move one →

Levels from Feng, McDonald and Zhang, “Levels of Autonomy for AI Agents”, Knight First Amendment Institute at Columbia University, 2025. The gate counts are ours.

That control center, for real

The control center your team logs into.

console.clustercode.ioDispatch

Dispatch, the guided home: asking for a DevBox, then shipping a feature through a workflow and opening the workflow run

Say what you want. Nucleus sets up the DevBox, runs the workflow, and asks before anything risky. Read the docs →

What that board does to your week

Your Tuesday, with the team running AI as a team.

Not a feature list. One day, hour by hour, from the seat of the person accountable for what ships.

07:10

Before standup

Nine tickets ran while nobody was here.

The 02:00 sweep took everything in "Ready for Dev". Each one got its own disposable DevBox on the office Mac minis. Six are green. Two are waiting at your gate. One failed its tests and looped twice before giving up, with the transcript to show why.

9runs overnight0humans awake

#482Rate limit the invite endpointAt your gate

#488Migrate notification kinds to the registryAt your gate

#470Fix stale plan badge after downgradeMerged 02:14

#479Worker heartbeat drops on sleepFailed · 2 loops

09:30

Review

Two at the gate. Ten minutes each.

A draft PR, the verification results, a second model's review notes, the cost. You read the summary, skim the diff, approve one and send one back with a sentence. No chat windows. No "what did you try".

console.clustercode.io · runsA run parked at the approval gate

A ClusterCode run paused with a question for the human, waiting at the approval gate

11:45

The one that went sideways

You take the terminal. Then give it back.

#479 kept failing on a flaky test. You open the live terminal on the same DevBox the agent used, see the timing issue in thirty seconds, fix the fixture, and hand the run back. The agent finishes it. The DevBox is thrown away after.

console.clustercode.io · displayThe live DevBox desktop

A live DevBox desktop in the ClusterCode console, open for a human to take over

14:00

The question from finance

"What did that run cost us?"

You open it and read the answer off the run itself: the AI and the compute it burned. Behind it sits every run the team has made, by project and by agent, with what each one took. The history nobody could have produced a month ago, because the runs lived in twelve private terminals.

console.clustercode.io · runsRun history, per project

The ClusterCode run timeline: every run, its agent, its duration and outcome

17:30

Leaving

You set tonight’s sweep and go home.

Eleven tickets are in "Ready for Dev". The same Schedule runs again at 02:00, on the same machines, with the same gate waiting for you at 07:10. The team did not have to agree on anything. The workflow is the agreement.

console.clustercode.io · schedulesTonight’s schedule

The ClusterCode schedules page with the nightly sweep armed

Pricing

Start free. Automate when you’re ready.

7-day free trial with full Pro access. No credit card required.

ClusterCode is in pilot. We're granting extended trials with compute to a limited number of teams.

Free

$0/mo

Bring your own machines and your own Claude subscription.

Get started free

✓3 self-hosted workers · 6 self-hosted DevBoxes
✓1 concurrent run · 1 project · 1 cron schedule
✓AI with your own Claude subscription
✓1-day run history

Starter

$9.99/mo

For solo developers getting started with AI coding agents.

Select Plan →

✓8 self-hosted workers · 16 self-hosted DevBoxes
✓5 cloud-worker DevBoxes (pay per minute)
✓Managed AI catalog (Claude, Copilot, Codex)
✓GitHub Issues integration · live web terminal
✓3 concurrent run · 3 projects · 3 cron schedules
✓$2/mo AI + $2/mo compute allowance
✓14-day run history

Give the team one workflow.

Start your free trial →

No credit card · 7-day full Pro access · Cancel anytime