CODRON
로그인
영문만 제공

이 페이지는 아직 영문만 제공됩니다. 나머지 화면은 한국어로 보실 수 있습니다.

01 · SETUP

Getting started

The desktop app walks you through this on first launch. Here is the same flow in text, for reference and for readers deciding whether to install.

  1. STEP 01

    Install a coding agent CLI

    Codron runs the coding agent you already use: Claude Code or OpenAI Codex, on your own subscription. Install and sign in to at least one on each computer you want to run work on: Claude Code or Codex.

  2. STEP 02

    Install Codron

    Download the installer for Mac or Windows and open it.

  3. STEP 03

    Sign in

    Sign in with Google or GitHub. Projects, runs, and decisions sync to your Codron account, so every computer you sign in on shows the same dashboard. There is nothing to host yourself.

  4. STEP 04

    Add a project and queue your first request

    Hit New project, pick a source (a fresh folder, a GitHub or GitLab repository, or an existing local folder), set the scope (how far the agent pushes before it parks), and describe what to work on. The run streams in real time while the agent works in the background. Not sure it's a project yet? Talk it through in Drafts, then hit Promote to project and the form fills itself from the conversation.

02 · LOOP

Request. Run. Decide. Review.

Describe the work, let the agent run it, answer decisions when needed, and inspect the result.

REQUEST1/4

You describe the work

A project holds the goal. Each request is one specific instruction in its queue.

RUN2/4

The agent runs it

Reads code, edits, tests, commits. A run is one execution; a request can span several runs.

DECISION3/4

It escalates when stuck

Only for questions only you can answer, and options come with the card.

RESULT4/4

It leaves a result

What changed, what was checked, reported tokens and cost. Queued requests run, then the project parks.

REQUESTS AND MISSIONS

Queued requests run and then park. To let the agent keep building on its own, press Continue mission: runs continue toward the project's ambition (prototype, product, or service), and Park after this run is the soft stop. When the ambition's goal is reached, the agent asks: raise the tier or park it here? It never raises its own tier. This is how unattended runs stay bounded: the loop won't rack up spend past a milestone you haven't approved.

03 · MACHINES

Every computer, one dashboard

Install Codron on each computer you work from. The dashboard shows every machine and the projects on it, and a run always happens on the computer that holds the project's checkout.

codron · settings · machines

MACHINES

mac-studio
local · gpu
4 projonline
linux-tower
local · headless
2 projonline
hetzner-eu
cloud · 24/7
3 projonline
m4-air
portable
1 projaway

Runs where the code is

Run and Resume start work on this computer's local checkout, with its secrets, tools, and GPU. Codron never quietly picks another machine for you.

Move work with Git

To continue on another computer, pause, commit and push, then pull there and run. CLI sessions, uncommitted files, and local variables stay on the machine that made them.

04 · REQUESTS

Write it once. Watch it run.

A request is one specific instruction: fix this, add that, check this. The agent plans, edits files, runs tests, and commits. Watch the run stream, or close the laptop and read the result later.

new project

PROJECT NAME

outbound-mailer

ENGINE

claude code · opus-5.5

SCOPE

prototypeproductservice

FIRST TASK

pull enriched leads from the crm, dedupe by domain, and stage an outbound batch. keep everything idempotent so a re-run never double-mails.

Create project & start

Live activity stream

Tool calls, edits, tests, all timestamped. Look for the running dot; leave the tab and it keeps going.

Result

Every run leaves a result when it stops: what changed, what was checked, reported tokens, cost, and engine. That's what you scan on the way to bed.

Input when needed

If a question comes up mid-run, the project flips to needs input and holds. See the next section.

05 · DECISIONS

Only when it needs your call

The agent escalates only for questions it can't answer without you. Three shapes cover everything: a plain decision (pick an option), an env_request (a secret it doesn't have), and a file_request (an asset it needs).

codron · agent workspace2 machines online

Overview

4 projects
DECISIONS1 pending
needs your calloutbound-mailer

prototype goal reached, raise the tier or park it here?

PROJECTS
hn-radar

cluster show hn launches by vertical

working
outbound-mailer

prototype goal reached, awaiting your call

needs input
stripe-cli-agent

regenerate webhook fixtures for 2026-08-01 api

working
codron-site

shipped /releases feed + changelog page

idle
USAGE361k tok · 30d
BY ENGINE
claudecodex

Interactive: click one of the two options in the decision card to resolve it and watch the project row flip.

decision

storage: sqlite or postgres?

sqlitepostgres

Standard pick-an-option. Options come from the agent, sometimes 2, sometimes N. A terminal option (e.g. park it here) stops the project instead of resuming it.

env_request

STRIPE_PRODUCT_ID

added, continue

Agent hit a missing secret. Card includes the why + a how-to link. You paste the value into the project's Variables tab (or the machine keyring), then approve to resume.

file_request

logo.svg for the site header

added, continue

Agent needs an asset it can't generate. Drop the file into the project's dropzone, then approve. Same shape as env_request, different payload.

TELEGRAM APPROVALS

Connect Telegram in Settings and every decision also arrives in your chat. One tap answers it from bed; the desktop card updates itself, and the run resumes. Same for env_request / file_request cards.

06 · CHAT

Two lanes for thinking out loud

Not every idea is a project yet, and even inside a project, some conversations aren't requests yet. Drafts hold ideas before they become projects; project chats stay with the work.

DRAFTS

Idea chats in Drafts, where you toss something around before it earns a project. When one grows legs, hit Promote to project and the new-project form fills itself: name, goal, reference docs lifted from what you pasted. Review, tweak, create.

draft · saas-app-ideas

what if the export was a webhook, not a download

pros: streams, no local file. cons: needs a receiver url.

yeah ok try it as a new project

Promote to project

PROJECT CHATS

Every project has chat sessions for thinking beside the work. When the discussion lands on something actionable, Draft request turns it into a request, intent plus action items, that you edit and add to the queue. No copy-pasting your own chat into a new prompt.

outbound-mailer · session · deliverability

our bounce rate is up. spf/dkim?

spf passes. dkim signature is wrong on the ses fallback path.

fix and add a check to ci

Draft request

07 · PREVIEW

Click the pixel. Codron finds the file.

Explaining a UI fix in prose means screenshots and guessing. In Codron's live preview, the element you click resolves to its source file and line. Add a memo, and the snap queues as a request for the next run.

preview · localhost:5173PICK ELEMENT
acme-metricsOverviewReports
MonthRevenueGrowth
Apr 2026$12,480+8.2%
May 2026$14,920+19.6%
Jun 2026$15,134+1.4%
Export CSV
SNAP

src/reports/ExportButton.tsx:42 · <ExportButton>

The app is running

Codron spawns your project's dev server and embeds it. Interact with it like any browser window: click, scroll, type.

Snap resolves to source

The pick tool crawls DOM + source maps + the framework's data-* attributes to find the actual JSX for the clicked element.

Queues as a request

A snap is a request linked to the selected element. It rides into the next run with your memo, so the agent sees where and why.

08 · DEV SERVER

One URL for the running app

Codron manages your project's dev server: spawns it, tracks the port, restarts it when a broken run kills it, and exposes it over a shared tunnel so your phone (or a teammate) can hit the same instance.

outbound-mailer · dev
frontend:3000
ready
backend:8787
ready
TUNNELhttps://outbound-mailer.ngrok.app

Public URL routed to :3000 · phone-safe · dies with the project.

Rescues itself

If a bad edit crashes the server mid-run, Codron notices, restarts, and the agent picks up where it left off instead of failing the run.

Tunnel per project

Optional, through your own ngrok account. Open the preview on your phone, share the URL with a teammate for a quick look, or point a webhook at it during development.

09 · VARIABLES

Secrets stay on your machine

Edit each project's .env in the app, Vercel-style. When a run needs a key it doesn't have, the agent asks by name with a how-to link. Values never sync to the cloud. Supabase stores project metadata, not your secrets.

outbound-mailer · variablesENV REQUEST
env vars needed

STRIPE_PRODUCT_ID

checkout can't run without the live product id

how-to: Stripe dashboard → Products → copy id

.env · localadded, continue
STRIPE_PRODUCT_ID

Per-project .env

Each project has its own file, edited in the app or in your editor of choice. The agent reads it at run start.

Machine keyring

Seed a value once per box (STRIPE_SECRET_KEY, OPENAI_API_KEY) and every project on that machine reuses it. env_request cards for keyed values auto-populate.

Never in the cloud

Values are read locally by the agent process. Only the key names (what was asked for, whether it was answered) round-trip through Supabase.

10 · ENGINES

Pick an engine per project.

Claude Code or OpenAI Codex, with the model and effort you choose. The model list follows the CLI you have installed, so a new release shows up once the CLI has run. Smart routing, on by default, sends small requests to a lighter tier and keeps hard work on your primary engine.

outbound-mailer · engine
Claude Code
default · opus-5.5
OpenAI Codex
gpt-6-sol
SMART ROUTINGon

Simple requests use sonnet-5 · medium; moderate requests use sonnet-5 · high. Complex work uses your primary engine. Sub-agents a run starts use sonnet-5; the reviewer stays on the run's model.

One engine per project

Different projects, different engines. A Rust systems job on Claude Code, a marketing site on Codex. Swap anytime from settings.

Smart routing

Before a run, a short classifier sizes the queued requests. Simple ones go to the lighter tier (Sonnet 5 on Claude Code, GPT-6 Luna on Codex), complex work stays on your primary engine, and a failed or late answer falls back to the primary engine. Related requests share a session per topic and model, so a follow-up resumes with its context instead of starting cold. On Claude Code, the sub-agents a heavy run starts use the lighter tier too, while the reviewer stays on the run's model.

Models that keep up

The model and effort pickers read the catalog your installed CLI already carries. When Anthropic or OpenAI ships a model, it is in the dropdown after the CLI runs, with the version it needs spelled out; nothing in Codron has to update first.

Guardrails included

Daily spend caps and automatic pauses come with every engine. A runaway run stops itself before it drains a budget.

QUESTIONS / FEEDBACK

Email anything from typos to crashes. Every report shapes what we ship next.

contact@vivatechlab.com