TurkeyCode
Turkey = Turnkey

One prompt in.
A tested app out.

TurkeyCode is an open-source build orchestrator. It runs a coding agent through research, planning, build, QA and code review, one phase at a time. You describe the app. It does the rest, unattended.

npx turkeycode run "a habit tracker with streaks and reminders"

Needs Node 20+, Git, and an agent CLI you are logged in to. Claude Code is the default.

What happens when you hit enter

  1. 01

    Scope

    It restates what it thinks you want. You correct it until the spec is right.

  2. 02

    Plan

    Research, then a plan: as many build phases as the work needs.

  3. 03

    Build

    One focused agent session per phase, on its own branch.

  4. 04

    Prove

    Fresh agents test it. Blockers get fixed and retested. Then it merges.

Nothing merges past a gate it did not clear. The gates are plain code, so the model cannot talk its way through.

Case study: same model, with and without a spec

We asked Claude Code for a TI-84 style graphing calculator twice, one session each time. The only difference was the spec that TurkeyCode's research step wrote from the same one-line prompt, in under a minute.

Without the spec

One session, bare prompt

Time
8 min
Cost
$3.97
Code
909 lines
Tests
0

Looked the part. The graphing keys did nothing.

With TurkeyCode's spec

One session, same model

Time
37 min
Cost
$14.94
Code
2,893 lines
Tests
205

Graphs, tables and matrices all worked. It shipped with 3 visible bugs.

Then the full pipeline. TurkeyCode's own run from that prompt took 1 h 47 min and left 226 passing tests. Its QA found 4 blockers along the way and had them fixed before anything merged.

The spec is what made the build work. The gates caught what the spec could not.

One prompt and one run each, so read it as a case study and not a benchmark. The full run took about three times as long as the session with the spec, and TurkeyCode does not yet record what a run costs.

Where it earns its keep: a backlog, unattended

The calculator is the easy case for a single session. Tickets are not. On one production codebase, TurkeyCode worked a backlog over 24 days: plan, build, review and push a branch for each ticket, without stopping to ask.

86tickets attempted, in 42 hours of run time
78finished as branches ready for a person to review
38 of 62reviewed tickets where its own review found something to fix

Where a ticket left a question open, it wrote down the default it chose and kept going. Fix sessions turned down 5 review findings as out of scope or needing a person. It is not a rubber stamp and it is not perfect either: 9 tickets finished with findings still open, 6 failed outright, and a later independent review found blockers in 4 of the 53 it had passed clean. It pushes branches. A person still merges them.

The short list

Free and open

MIT licensed. Runs on your machine with your own agent login. No account here.

Any stack

Web, CLI, mobile, games, libraries. QA picks its tools from the project it finds.

Any agent

Claude Code, Codex CLI, Gemini CLI, OpenCode, or your own, with one flag.