convez.ai
AgenteDesenvolvimento WebIntermediário
ChurroStack

Churro Coder

ChurroStack/churro-coder

Agent de código local com fluxo PR nativo e multi-modelo

Ver no GitHub
BomPluginScore · de 0 a 100

Agent de codificação open-source que executa localmente sem cloud. Oferece fluxo completo de plano → código → revisão → PR com suporte a múltiplos modelos (Claude, Codex) em uma única interface, com execução sandboxed e rastreamento de custos integrado.

Instalação

$npx convez-ai ChurroStack/churro-coder

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
4
Forks
1
Acompanhando
0
Saúde
50%
Último commit
há 3 meses
Versão
v0.1.13
Issues abertas
1
Contribuidores
3

Casos de uso

  • Desenvolvimento de features com revisão automática de IA
  • Workflow colaborativo arquiteto (plano) + implementador (código)
  • Projetos sensíveis que não podem usar cloud ou sync remoto

Veredito

Ferramenta ambiciosa que reimplanta a experiência de agents de código (tipo Cursor) para desktop multiplataforma com sandbox real e orquestração multi-modelo. Produção-ready, mas ainda sem code-signing nativo — ideal para equipes que priorizam controle local sobre conveniência cloud.

README

A tradução para o português deste README ainda está na fila.

Churro Coder

Open-source, fully offline coding agent client. Run Claude Code, Codex, and more — locally, no cloud account needed.

Churro Coder is a desktop app for working with AI coding agents (Claude Code, Codex, …) on your own machine. Sub-chats, terminals, file viewers, plans, and diffs are first-class dock panels you can split and rearrange; each chat runs in its own git worktree so the agent can't trample your main branch; everything stays on-device — no login, no cloud sync, no analytics.

Cursor-like UI demo

Highlights

  • Plan → Code → Review → PR workflow. A first-class stepper for the way most engineers actually ship: the agent drafts a plan, you approve, it codes in an isolated worktree, runs an AI review on the diff, and opens the PR — each phase has its own panel and one-click actions.
  • Mix Claude and Codex in a single chat. Route each phase to the model that's best at it — e.g. Claude Opus as architect to design the plan, Codex/GPT to implement it, Claude Opus again to review the outcome. No copy-pasting between tools; one chat orchestrates them and you get the best of every provider.
  • Sandboxed execution. Agent file access is locked to the chat's worktree, your config dirs (~/.claude, ~/.codex, ~/.churrostack), standard scratch space, and optionally toolchain caches — nothing else. On macOS and Linux the sandbox is enforced at the OS level (Seatbelt / bubblewrap); on other platforms the SDK-level path enforcement blocks reads and writes outside the allowed roots. Configurable per-project, per-chat, or globally from Settings → Sandbox.
  • Spend tracking built in. See total usage and per-chat cost broken down by provider and model (usage.png), so the multi-model workflow above doesn't turn into a billing surprise.

Downloads & releases

Grab the latest installer for your platform from the GitHub Releases page:

PlatformWhat you download
macOS.dmg installer (and a .zip of the app), for both Apple Silicon (arm64) and Intel (x64)
WindowsNSIS installer .exe (choose install dir) or a standalone portable .exe
Linux.AppImage (run-anywhere) or .deb package

The macOS and Windows builds are not code-signed yet. They run fine, but the OS will warn you: on macOS, right-click the app → Open the first time (Gatekeeper); on Windows, click More info → Run anyway (SmartScreen). Linux AppImages just need chmod +x.

Releases are automated with release-please: merged Conventional Commits accumulate into a "release PR"; merging that PR tags the version and triggers .github/workflows/release.yml, which builds all three platforms on their native runners and attaches the installers to the Release. Installers are only built when a release is cut, never on ordinary PRs. Deeper mechanics live in apps/desktop/docs/release.md.

Enabling macOS code signing & notarization

The release workflow is already wired for signing + notarization but ships disabled until an Apple Developer ID certificate is available. To turn it on, add the following repository secrets under Settings → Secrets and variables → Actions → New repository secret — no workflow edit is needed; the next release will automatically sign and notarize the macOS build:

SecretWhat it holdsNeeded for
CSC_LINKbase64 of the exported Developer ID .p12 certificate — base64 -i cert.p12 | pbcopysigning
CSC_KEY_PASSWORDthe password you set when exporting the .p12signing
APPLE_IDENTITYDeveloper ID Application: Your Name (TEAMID)signing
APPLE_IDyour Apple ID emailnotarization
APPLE_APP_SPECIFIC_PASSWORDan app-specific password from appleid.apple.comnotarization
APPLE_TEAM_IDyour 10-character Apple Team IDnotarization

The workflow treats CSC_LINK as the switch: present → it signs the macOS build with the cert; absent → it forces a clean unsigned build. These secrets are scoped to the macOS runner only, and notarization runs only when the three APPLE_* notarization secrets are all set. (The same variables are listed in apps/desktop/.env.example for local signed builds.) Windows and Linux currently ship unsigned; wiring Windows Authenticode signing is a separate, not-yet-done step.

Repository layout

This is an Nx monorepo with three apps that share a single product surface but speak different stacks on purpose:

AppStackWhat it does
apps/desktopElectron + Vite + React + TypeScript + Tailwind, bun-managedThe local UI on macOS / Windows / Linux. Hosts the agents, the dockview workspace, the SQLite store, and bundled CLI binaries.
apps/gatewayASP.NET Core (.NET 10)Cloud-side NAT gateway between a daemon and clients (used for remote / browser access).
apps/daemonGo (cross-platform CLI)Installable daemon that the gateway talks to on behalf of remote clients.

The desktop app is the primary surface; the gateway + daemon are how Churro Coder reaches a machine you're not sitting in front of.

For app-level architecture, build pipelines, schema, and gotchas, read each app's AGENTS.md:

  • apps/desktop/AGENTS.md — windowing, tRPC routers, Drizzle schema, release pipeline
  • AGENTS.md — repo-level conventions for AI coding agents working in the workspace (CLAUDE.md is a symlink to it)

Quick start

Prerequisites

  • Node.js ≥ 20
  • pnpm (the version pinned in package.json#packageManager — install via corepack enable)
  • Bun (install) — apps/desktop is bun-managed
  • .NET 10 SDK (only if you'll run gateway)
  • Go ≥ 1.26 (only if you'll run daemon)
  • Xcode Command Line Tools on macOS, build-essential on Linux — needed to compile native node modules (better-sqlite3, node-pty)

Install + run

# Root: pnpm drives Nx + plugins
pnpm install

# apps/desktop is intentionally outside the pnpm workspace and uses bun
cd apps/desktop && bun install && cd -

# See every Nx project + the targets each one exposes
pnpm exec nx show projects        # → ["desktop","gateway","daemon"]
pnpm exec nx graph                # interactive dependency graph

# Run any single app
pnpm exec nx run desktop:dev      # Electron app with HMR
pnpm exec nx run gateway:run      # .NET API on http://localhost:5111
pnpm exec nx run daemon:serve     # Go daemon

# Or build everything (Nx caches; second run is instant)
pnpm exec nx run-many -t build

For platform installers (DMG / NSIS / AppImage / DEB) and the full release pipeline, see apps/desktop/README.md and apps/desktop/AGENTS.md.

A note on the two package managers

pnpm lives at the repo root and drives Nx + plugins. Bun manages apps/desktop exclusively — pnpm-workspace.yaml deliberately excludes that path. Don't run pnpm install inside apps/desktop; the two package managers will fight over the lockfile. This split is intentional, not a migration in progress.

Testing

# Run all test suites (Vitest + xUnit + go test), Nx-cached
pnpm exec nx run-many -t test

# Run only tests affected by your local changes
pnpm exec nx affected -t test
AppRunnerHow to run directly
apps/desktopVitestcd apps/desktop && bun run test
apps/gatewayxUnit (WebApplicationFactory)dotnet test apps/gateway/Gateway.Tests/Gateway.Tests.csproj
apps/daemongo testcd apps/daemon && go test ./...

Desktop tests run in Node.js (no Electron). Tests that need a DOM add // @vitest-environment jsdom at the top of the file. See AGENTS.md#running-tests for more details.

To generate an HTML coverage report for the desktop suite:

cd apps/desktop && bun run test --coverage && open coverage/index.html

CI

GitHub Actions runs nx run-many -t build lint test on every PR to main (see .github/workflows/ci.yml). Wire the CI / Build & check check into a branch protection rule / ruleset to gate merges.

Contributing

Issues and PRs welcome. The repo is set up for AI coding agents too — AGENTS.md (and the symlinked CLAUDE.md) is the single source of truth for any tool working in this codebase. If you're adding a new app, follow the patterns in AGENTS.md#adding-a-new-project.

License

GNU Affero General Public License v3.0 — see LICENSE for the full text.


Originally based on 21st-dev/1code.

Ver o original no GitHub

Bom para

Relacionados