Go to bed with a plan. Wake up to verified commits.

Nightshift runs Claude Code and Codex overnight, in parallel, and lands only the work that passes your tests on the real branch. No pile of pull requests to triage in the morning. No agent's word taken for "done".

Run it on your machine Read the source

Open source · Apache-2.0 · Runs locally · Claude Code and Codex

A night's run, watched live in the Studio.

Writing code got cheap. Checking it didn't.

Agents can produce a week of changes in a night. Then you spend the next day finding out which of them are real: the tests that were quietly deleted, the fix that works only on a stale branch, the confident summary of work that never passed. More agents just make that queue longer.

Nightshift moves the checking to before anything lands, so the morning starts with results.

Agents don't mark their own homework

Every change is built and tested in a clean checkout of the real branch head, after everything that landed before it. The agent's claim is recorded. Your tests decide.

A second opinion from a different model

Risky work is examined by another provider's model before it lands, because one model reviewing itself shares its own blind spots. When they disagree, a ruling settles it, on the record.

Frontier spend only where it matters

Small, well-tested jobs go to cheap models and climb only when they fail. A budget stops new work before an overnight run turns into a surprise bill.

Know why every line is there

Each decision is kept with the options it beat and the commits it produced. Disagree with one in the morning, reverse it, and the fix is planned from the record.

Decide the expensive things. Delegate the rest.

You plan with Nightshift until only cheap choices are left, and every credential or human step is done up front. The run doesn't stop at 3 a.m. to ask you something.

Nothing touches your branch but Nightshift

Agents work in isolated worktrees and never commit. Nightshift owns every commit and lands them one at a time, so parallel work can't collide on your branch.

One evening of planning, one night of work

Plan it together. A short plan and a contract: the outcomes, the seams, the decisions that would be costly to get wrong.
Ratify it. A deterministic check says when it's ready to run unattended.
Let it run. Parallel agents, each change verified before it lands. Watch it live in the Studio, or go to sleep.
Read the report. What landed, what's waiting on you and why, what it cost, and every decision behind it.

What it doesn't do

It won't make a bad test suite trustworthy. Nightshift is only as strict as your checks.
It doesn't host your agents yet. Runs use your machine and your own Claude or Codex subscription.
It doesn't replace reading the code that matters. It makes sure that what you read has already passed.

Run it tonight

Node 24, git, and Claude Code or Codex signed in. No cloud account.

git clone https://github.com/wildorder/nightshift.git && cd nightshift
npm ci && npm run build
(cd apps/cli && npm link)

nightshift local        # the control plane and the Studio, on this machine

Then nightshift init in a repository of yours, plan with /plan-program in Claude Code, and nightshift run. The README has the rest.

Built in the open, the way it runs

Every program that built Nightshift has its contract, its decisions and what actually got built in the repository. Start with the vision, the architecture and the programs.