→AI coding agent · Open source

An autonomous coding agent that runs inside your GitHub, not ours.

File an issue, apply one label, and get a pull request that a human still has to merge. Everything executes in your own GitHub Actions, on your runners, with your secrets. SimplyCubed hosts nothing and never sees your code.

Status

Beta. The architecture is documented and readable today. There is no release to install yet, so the way in is to watch the repo. If you need an agent working against real code this quarter, start with the readiness work below.

Read the source on GitHub

30 minutes. Your repos, your CI, and what’s missing before an agent touches them.

Apache-2.0 · your runners · a human merges every pull request

open source license: Apache-2.0
lines of your code reach us: 0
label (sc:go) is the only trigger: 1
of merges are human, by design: 100%

01The blocker

The blocker on AI coding agents is not capability. It is what they can reach.

An agent that can merge has unbounded blast radius

Give a loop write access to your default branch and a bad change is already in your history by the time anyone reads it. Review after the fact is not a control.

An ungated loop wanders

With nothing to stop it, an agent churns, breaks things, and still reports success. The check that would have caught it usually does not exist yet.

Evaluations die in security review

Whether your source and your secrets leave your perimeter is the question that ends most pilots. It is also the right question to ask.

So put the agent inside your GitHub instead of sending your code out of it.

02What it is

Issue in. Reviewed pull request out. A human merges.

SimplyCubed Code is designed to turn a GitHub issue into a reviewed pull request with no person writing the code and a person always deciding. A human files the issue. The agent implements it on a branch, a separate read-only reviewer role reads the diff, a fixer role addresses the findings, and the pull request goes out for a human to merge.

It proposes. It does not dispose. It never pushes to your default branch and never merges its own work. Configuration is one file .github/simplycubed.yml.

You apply one label. The bot owns the other five. Only a person merges.

sc:goStart work on this issue. The only trigger a person sets.Set by you
sc:queuedAccepted, waiting to start.Set by bot
sc:workingImplementing the change on a branch.Set by bot
sc:blockedNeeds a human. The agent stopped and left a note.Set by bot
sc:reviewReview and fix pass in progress.Set by bot
A human reviews the pull request and merges it. Only then does the issue close.
sc:doneThe bot closes out the issue.Set by bot

An issue sits in one state at a time. The prefix is configurable with labelPrefix, and defaults to sc.

03How it works

One label in. A pull request out.

Step 101
File and label
A person opens an issue describing the change and applies sc:go. That is the whole trigger.
Step 202
Implement on a branch
The agent picks the issue up, works on a branch, and runs your gate against what it wrote.
Step 303
Review and fix
A separate read-only reviewer role reads the diff. A fixer role addresses the findings and runs the gate again.
Step 404
The gate passes
Nothing is called done until your own quality check is green. sc:blocked is a first-class outcome: the agent stops and leaves a note rather than pushing through.
Step 505
A human merges
The pull request opens for a person to read. The agent stops here, every time.

04Where it runs

Your runners. Your secrets. Your branch protection.

The strongest action available to this agent is opening a pull request against a branch. It holds no deploy credentials, it never merges, and a human decides what lands. Everything below is what that means concretely.

No server on our side. The whole thing is event-driven GitHub Actions. There is no server and no VM for SimplyCubed to operate, so when you install the app and file issues, the work is designed to execute on your runners inside your organization.

Your code stays in your repos. SimplyCubed never receives it. There is no ingest, no index, and no copy on our side, because there is no our side.

Your secrets stay in your secret store. Your model provider keys, your GITHUB_TOKEN, and anything else are read by your own Actions runs.

Your minutes, your bill. You pay GitHub for the compute and your model provider for the tokens. The open source project sends nothing to us.

One audit signal. The GitHub App identity is simplycubed-code[bot], and it is the single audit signal for everything the agent does.

Five questions to ask of any coding agent

Can it merge its own work?: No. A person clicks merge, every time.
Does it hold deploy credentials?: No. The strongest action available to it is opening a pull request.
Is a passing gate mandatory?: Yes. A repo with no gate is refused.
Can you read the source?: Yes. Apache-2.0, public, today.
Whose model provider account?: Yours. You bring the key and you pay for the tokens.

The GitHub App permissions, exactly

contents:write

Read repository contents and write to a branch. This is how a branch and its commits get created.

issues:write

Read issues, comment on them, and set the lifecycle labels.

pull_requests:write

Open a pull request and comment on it.

metadata:read

Read basic repository metadata. GitHub requires this of every app.

What it can never do

Merge a pull request. Not its own, not anyone else’s.
Push to your default branch, or bypass your branch protection rules.
Hold deploy or production credentials. There is no path to production.
Send your source or your secrets to SimplyCubed.

Verify the permissions on the GitHub App page → (opens in a new tab)

05Built in the open

Read it before you trust it.

Apache-2.0, public today. Every claim on this page is in the source, so you can check it instead of taking our word for it. These are the files worth opening first.

STATUS.mdWhere the project actually stands loop/goal.jsonThe machine-gradable milestone contract docs/decisionsArchitecture decisions, with the reasoning docs/setup.mdHow installation and configuration work MakefileThe gate the project runs on itself .github/workflows/check.ymlThe required CI check LICENSEApache-2.0, in full SECURITY.mdHow to report a vulnerability

Agent-authored PRs

Every pull request the agent opens in its own repository will be listed here, under the simplycubed-code[bot] identity. The list is empty. That is the number to watch.

See the agent’s pull requests → (opens in a new tab)

06Status and roadmap

In beta. No releases yet, no dates invented.

The repository is public and in beta. There is no release to install yet, and the roadmap below is order, not schedule.

Roadmap, roughly in order

01The core issue-to-pull-request loop with the reviewer and fixer roles.
02The self-onboarding flow via bootstrap issue and setup pull requests.
03The Codex on Azure OpenAI engine adapter.
04The Claude Code engine adapter.

You own your CI

Run it yourself

Apache-2.0, free, no SimplyCubed account.
You need Actions, your own model key, and a gate you trust.
Nothing is installable yet. Watch the repo.

What running it yourself takes →

You need someone to own it

Have us make your org agent-ready

A real gate per repo, each one seen to go red.
Branch protection, scoped permissions, environment separation.
Deliverable now, and useful with any coding agent.

See what that covers →

Or score your org first, in 3 minutes →

07Run it yourself

What running it yourself takes.

Free, open source, no SimplyCubed account, and nothing to buy. Onboarding is designed to happen through GitHub itself rather than a local setup script.

Prerequisites

GitHub Actions on the repo, with minutes to spend.
Your own model provider key. The Codex CLI on Azure OpenAI adapter is first, and a Claude Code adapter is planned.
A gate you trust: one command that fails when the code is wrong.

.github/simplycubed.ymlyaml

labelPrefix: sc

gate: make check

The gate command is the required key. Everything else has a default.

The sequence, once it is installable

01Watch the repo. Nothing is installable yet, and the install link will appear there first.
02Install the GitHub App on one repo, once that is available.
03File the bootstrap issue.
04Read the setup pull requests the agent opens and merge what you agree with. Setup arrives as pull requests, so you see it before any of it lands.
05Add .github/simplycubed.yml with a gate command.
06File your first real issue and apply sc:go.

If nobody on your team can own the gate work, that is the engagement in the next section.

08Have us roll it out

We make your GitHub org safe for any coding agent, starting with the check that has to fail.

This work stands on its own with any coding agent, including the ones you already pay for.

Agent loops fail in real codebases at the point where nothing checks the work. Most teams cannot adopt any coding agent yet because the check that would catch a bad change either does not exist or has never been seen to fail.

Make your GitHub org agent-ready

A scoped first engagement, one senior engineer who owns it end to end, and direct access to the founder.

The gate is the engineering. The loop is the easy part.

A real gate per repo. Typecheck, tests, and a build wired as the required check, each one broken on purpose first to prove it goes red. A check only ever seen green cannot stop a bad change.
Branch protection with no bypass, required review, and bot identities scoped to what they need, so no automation can put code in main without a person.
Environment and credential separation, so no agent identity can reach production.
Guardrails on the merge path: destructive-migration gates, one-command rollback, and a post-deploy check over the real customer path.
A pilot on 1–2 repos, then rollout across the org.
Handover. Documentation and walkthroughs, running in your CI, owned by your team.

This runs as an AI Automation Sprint. The first one delivers the agent-ready foundation on your pilot repos: a gate that has been seen to go red, branch protection with no bypass, and identities that cannot reach production. Organizations with more repos or more environments than one sprint covers run the same play again for the next set, scoped on the call.

Sprint one: $10K. Each additional surface is scoped in the same $10K–$20K band, agreed upfront.

You get the gate and the guardrails whether or not you ever run our agent. When SimplyCubed Code is installable, rollout is a next step instead of a project.

Success criteria, agreed in writing before we start

A required check on every in-scope repo, each one demonstrated to fail.
Zero paths to main without a human.
No agent identity holding a deploy credential.

Each one is binary: it is true on your repos or it is not. Same guarantee as the AI Automation Sprint on the homepage. If we do not hit them within 30 days of deployment, we keep working at no extra cost until we do, or we refund 50%.

This is the same standard SimplyCubed runs on every repo it owns, built by a founder who ran security for finance and payments companies. Least-privilege access and audit are where the design starts.

30 minutes. Your repos, your CI, and what’s missing before an agent touches them.

Not sure where you stand? Get your Agent-Readiness Score first.

→Why us

Loops and gates are how SimplyCubed builds, before they were ever something to sell. Every repo it owns ships with a required gate, a bot identity that cannot deploy, and a human on the merge. The founder ran security for finance and payments companies, so least-privilege and audit are where the design starts.

About the founder → (opens in a new tab)

Recognized by Clutch as a top AI company in Japan, 2026.

09Fit check

Is this for you?

Worth your time if…

You have a backlog of small, well-specified changes and nobody free to do them.
Your source cannot leave your perimeter, and that has ended an agent evaluation before.
You are on GitHub, with Actions and branch protection you control.
You have a CI check that has been seen to fail, or you are willing to build one first.
You would rather read an agent’s source than take a vendor’s word for what it does.

Not for you if…

You need SimplyCubed Code itself running against real work this quarter.
You want a hosted service that works with no engineering on your side.
You are not on GitHub.
You want the agent to merge its own work. It will not.
You are shopping for a coding assistant in your editor. That is a different thing.

10Questions

The hard questions

Is it installable yet?

Not yet. It is in beta with no release, so there is nothing to point at one of your repos this week. The roadmap above is in order and without dates, because order is what the project can commit to. Watch the repo for the release. The gate work it depends on is worth doing before then either way.

What does it cost?

The software is open source under Apache-2.0 and free. Running it costs your GitHub Actions minutes and your model provider tokens, billed to you by GitHub and your provider. Having SimplyCubed make your organization agent-ready is a paid engagement: $10K–$20K per sprint, $10K for the first one, agreed upfront. Most orgs start with one sprint.

Can it merge or deploy?

No. The GitHub App is granted contents:write, issues:write, pull_requests:write, and metadata:read. It holds no deploy or production credentials. The strongest action available to it is opening a pull request against a branch, and your branch protection rules and a human decide what happens after that.

Do you see our code?

No. The design runs everything in your own GitHub Actions, on your runners, with your secrets. SimplyCubed operates no server in that path and receives neither your source nor your credentials. That claim is in the source, under Apache-2.0, for you to check.

Which models does it use?

Whichever you bring. The model sits behind a pluggable Runner interface. The first adapter targets the Codex CLI against Azure OpenAI (GPT-5.4 and 5.4-mini), and a Claude Code adapter is planned.

Why is a gate mandatory?

Because an agent loop with nothing to stop it will churn and still report success. The gate is the safety mechanism, so a repo without one is refused rather than run against a default that papers over the gap.

What if our CI is not worth trusting?

Then start there. A check that has been seen to go red is the prerequisite for any coding agent, not only this one, and it is the engagement SimplyCubed sells today. A check only ever observed green is indistinguishable from no check.

Can we contribute?

Read, audit, use, and fork freely under Apache-2.0. External code contributions and pull requests are not currently accepted. Bug reports and security reports are welcome through issues and SECURITY.md.

→Two ways in

Read the source, or get your repos ready for it.

The repository is open and in beta. The gate and guardrail work that makes any coding agent safe to accept is available now.

Not ready to talk? Score your org in 3 minutes →Watch the repo →