We build software with AI coding agents all day. So we built the tool we were missing, then built the system that gets it found. This is the story of VibeiDE, a desktop app that runs Claude Code, Codex and OpenCode in parallel across many projects, with every task waiting for review before it counts as done.
It is our own product, which makes it an honest case study: we were the client, the designers, the engineers and the operators, and we have the search data to show what happened after launch.
The problem: one terminal, too many requests
The idea came from a very ordinary failure. Patrik, our founder, was putting many requests into one agent terminal. The agent silently skipped some of them, and he forgot to check others. With several projects open at once, the day turned into switching between tabs and terminals and losing track of what had been asked where.
Long, mixed sessions had a second cost. They burned tokens and pulled unrelated context into answers, which caused mistakes that a fresh session would not have made.
Before building anything, we checked whether this was just us. Developer threads on Reddit in September 2026 described the same pain in their own words: "tab hell", agents skipping requests in long sessions, losing track across projects, and IDE terminals freezing once enough sessions were open.
The product: one task, one terminal, one fresh session
The core rule of VibeiDE is simple. One task is one terminal and one fresh session, and nothing counts as done until you review it. Everything else follows from that rule.
- Per-project task queues. Waiting prompts can be edited and reordered. A task starts when its provider has a free slot.
- Parallel limits per provider. Each project has its own slot meters, so you decide how many Claude Code or Codex sessions may run side by side.
- Ready for review. Finished tasks land in one list instead of scrolling away in a terminal.
- Resumable conversations. Every task keeps its provider conversation, so you can pick it up later without re-explaining.
- A Git worktree per task. Each task can get its own worktree and branch, so tasks run side by side without editing the same files.
- Plan council. One agent drafts a plan, a second agent reviews it critically, the author revises, then you decide.
- Built-in diff view with commit, or commit and push.
- Usage display with provider-reported usage, runway and reset countdowns.
- Grid view or Office view, an isometric office where the agents work at desks.
project queue provider slots review
+----------------+ +------------------+ +------------------+
| task 4 (edit) | | Claude Code 2/3 | | task 1 diff ok |
| task 5 | ---> | Codex 1/2 | --> | task 2 diff ? |
| task 6 | | OpenCode 0/1 | | resume or commit |
+----------------+ +------------------+ +------------------+
one task = one terminal = one fresh session, own worktree if enabledFor teams, VibeiDE has a team dashboard and a manager office view. Privacy was a design input from the start: there is no per-person leaderboard, and the team view shows task metadata only.
Architecture: stay inside the provider's rules
The most important architecture decision was what VibeiDE does not do. AI providers are rightly strict about how their subscriptions are used, and a tool that works around them is a liability for its users.
So VibeiDE spawns only the official CLI binaries with the user's own login. It handles no OAuth tokens, uses no agent SDK, makes no API calls of its own and never reads credentials. Each user brings their own provider accounts, and VibeiDE never bypasses provider limits. It is an orchestrator around the tools people already pay for, not a replacement for them.
That decision shaped the runtime:
- Claude Code and Codex run in embedded terminals that show the original CLI, so the user sees exactly what the provider's tool shows.
- OpenCode runs headless, and VibeiDE records its activity for the review view.
- Billing runs on Stripe, with team seats managed inside the app and a separate sandbox account for testing.
- The marketing site ships in 11 languages: English, Czech, German, Spanish, French, Hungarian, Italian, Dutch, Polish, Portuguese and Slovak.
macOS is fully validated, Windows is supported and Linux is experimental. The source is available under the PolyForm Noncommercial 1.0.0 license, and commercial use needs a license. Pricing starts with a 14-day trial without a card, then $7.99 a month, $79 a year or $249 once, with a Team plan at $14.99 per seat per month.
This is the kind of work we do under desktop app development: a native-feeling app that wraps other processes safely, with billing and licensing that hold up in production.
How we built it: VibeiDE built VibeiDE
We dogfooded from the first usable build. VibeiDE managed our own projects in parallel, including VibeiDE itself and Namiru.ai, so every rough edge hit us before it hit a user.
One practice made the biggest difference: per-project long-term memory. Claude Code and Codex hooks read a project wiki at the start of each task and document the task at the end. Fresh sessions stay fresh, but they do not start from zero. The agent knows the conventions, the decisions already made and the traps already found.
The pace that this allowed is visible in the public record. The first commit landed on 16 July 2026. Version 0.1.0 shipped on 12 August 2026. By 6 October 2026 the VibeiDE changelog listed 163 versions, the latest being 0.4.1.
The landing page was designed with Claude Design and then built by hand into the production site.
Getting it found: the growth engine
A good product on a new domain is invisible. So the second half of the project was a growth system, built the same way as the product: with agents, on our own tooling, with a human approving anything public.
A custom SEO MCP server. We built an MCP server that connects Google Search Console, Bing Webmaster Tools, IndexNow, Google Keyword Planner, PageSpeed and Lighthouse, a site crawler, Common Crawl link-gap analysis, an outreach tracker and a decision log. Any agent can read the real data instead of guessing.
A daily autonomous SEO agent. Every day it reads that data, fixes indexing issues, researches keywords, writes and publishes articles with animated SVG diagrams, edits and deploys the site, submits URLs, and records every decision with a review date. Its limits are explicit: it never spends money and never posts outside the site.
A daily outreach task for the founder. One backlink target a day, with ready-to-paste text. Outreach is the one part we deliberately keep human.
Build in public on X. The @vibeide_dev account is run by a posting agent, and every post is approved by a person before it goes out.
Search Console Bing IndexNow Keyword Planner Lighthouse crawler
\ | | | | /
+------------------ SEO MCP server --------------------+
|
daily SEO agent (no spending, site only)
| | | |
fix indexing research write and log every
keywords deploy decisionResults so far
All numbers below come from Google Search Console, Bing Webmaster Tools and the site's own analytics, pulled on 6 October 2026.
Content. 20 in-depth articles were published in 8 days, from 29 September to 5 October 2026, each 1,000 to 1,500 words with three animated diagrams. The sitemap holds 338 URLs across 11 languages with zero errors in Google and Bing.
Indexing. On 4 October 2026, Google URL Inspection showed 46 of 52 English pages indexed (88%) and 232 of 342 URLs indexed across all languages.
Google Search, 6 September to 5 October 2026:
| Metric | Value |
|---|---|
| Impressions | 1,895 |
| Clicks | 16 |
| Average position | 10.2 |
| Impressions a day before the blog (7 to 28 Sept) | about 21 |
| Impressions a day after launch (1 to 4 Oct) | about 333, roughly 16x |
| Share of the month's impressions in the 6 days after launch | 75% |
The top page, a guide to installing the OpenAI Codex CLI, reached 448 impressions at an average position of 14.8 two weeks after publishing. Google data for 3 and 4 October is still preliminary.
Bing. All 342 URLs were submitted through IndexNow and all 12 sitemaps were accepted. There are no Bing impressions yet, which is normal for a new domain without backlinks.
What we are honest about
Two things are not there yet, and a case study that hides them would be a brochure.
- No blog-attributed signups yet. The blog cohort is too new and too small: 5 tracked readers in 30 days who gave analytics consent. Impressions are a leading indicator, not revenue.
- Backlinks are the open gap. Outreach started in October 2026. Search engines trust a new domain slowly, and no amount of on-site work replaces other sites linking to you.
What is next
The roadmap is clear: team features, more agent CLIs (Grok, Copilot, Cursor CLI and Kimi are on the list) and maturing Windows and Linux to the level macOS is at today.
What this means for your project
VibeiDE shows the full loop we offer clients: find a real problem, design a product around one clear rule, build it with AI-assisted engineering that stays reviewable, respect the platforms you depend on, and then build the system that gets it found.
If you need an AI-native product built and launched, from custom AI development to the desktop or web app around it, talk to us. And if you run several coding agents yourself, try VibeiDE free for 14 days.