Ship AI features in Spring Boot — this week.
Chat APIs, embeddings, RAG, agents — real working code, no theory bloat.
Get the book →The engineering studio for building exceptional software. One prompt in — a researched, planned, built, and tested project out.
Intel Mac?
Secure checkout via Razorpay — card or PayPal. The download link is only active for 10 minutes after payment, so stay on this page until you've grabbed it.
Want to try it first? Download the free trial for Mac
3 projects, any provider — no payment needed. What's included?
Mac says the app is "damaged" and won't open? That's expected for now — one Terminal command fixes it.
Windows says "Windows protected your PC"? Expected too — click More info → Run anyway.
Every agent, every pipeline stage, real git commits — the full engineering process, on 3 real projects.
Nothing about the pipeline itself is limited or watered down on any of them. Once all 3 are used, buy the full version to keep building — a separate, uncapped download.
A small engineering org, simulated end to end — every stage handing off to the next.
Research, Vision, Architect, Sprint Planner, Senior Developer, Tester — specialist agents hand off work through a real development lifecycle, not one giant prompt.
Nothing is built until you review the proposed backlog. Steer mid-run, pause and swap models, or regenerate the plan with feedback.
Anthropic, OpenAI, OpenRouter — or fully local with Ollama and LM Studio. Local models always cost $0, and your keys are encrypted with your OS's built-in secure storage.
The project's real build command (Maven, npm, Gradle, Go, Cargo, ...) must pass before a run is allowed to finish — failures route back to the developer automatically.
A true PTY shell is embedded right in the completion summary, alongside the exact commands to run and test what was built.
Cloud model pricing is checked live; local models are always $0 — never a guess, never a stale number quietly assumed to still be true.
Real screenshots from a real project built with Atlas Studio — a Spring Boot + Spring AI MCP server with one tool, from connecting a model to a working, tested build.
Anthropic, OpenAI, OpenRouter, Groq — or a fully local server like Ollama or LM Studio. Add as many as you want and switch per project. Local models always cost $0 and never leave your machine.
The full, in-app release notes are one click away after you install.
Version 2.3.0 went out without its Windows installer. 2.3.1 ships the same features with the Windows build included.
Tasks that work on separate files build side by side, each in its own git workspace, then their changes are merged in backlog order.
The sprint planner lists the files each task changes and the tasks it needs. Overlapping or dependent tasks still run one after another.
A task that changes files outside its plan, or clashes with another task, is marked failed instead of merged, and the reason is shown.
Version 2.2.0 went out without its Windows installer. 2.2.1 ships the same features with the Windows build included.
The composer has an Ask and a Build mode. Ask answers from the code; Build starts a new run on the project, with planning and your approval first.
Before the first build, Atlas Studio checks that the tools the project needs are installed, and names what to install if one is missing. Retry carries on.
Virtual environments, installed packages, and build output that a run created are removed when it ends, so the folder keeps only the code you asked for.
New and imported projects get a .gitignore covering Python, Node, Java, Go, Rust, .NET, and common services, so only source files get pushed.
Each column shows three projects first. Show all opens the rest, so Done no longer grows the page.
The model picker and the Ask/Build switch sit on one row, with the text box and Send below.
Send a bug or an idea from the sidebar or the website. It becomes a public issue you can follow, with only the app version and platform attached.
The overview on the Projects page shows projects, agent runs, tasks done, and tokens, with the input and output split under the total.
The Release Notes window opens on the current release; pick any older version in the row above it.
On Windows, the end-of-run app check now stops the app it started, so it no longer keeps running in the project folder.
Open a project in IDE mode: the complete run view (stages, approvals, continue) sits beside your files and editor, and the sidebar folds into the top bar.
Create, rename, copy, paste, upload, and delete files and folders, then save edits — once no stage is running.
Open as many terminal tabs as you need in a project, like a normal IDE.
Markdown files preview in the IDE, and Mermaid diagrams render in chat and in documents.
Point Atlas Studio at a folder with code: it maps the project, documents the architecture, and starts a build if you add a note.
Agents compile and test while they work; Atlas Studio starts the app only after the build is green, and reports if it does not run.
Each run writes a dated report to atlas/run-reports with tokens per stage, backlog status, problems, and follow-ups.
The summary shows an estimated cost range from published list prices, with the price sources found by web search.
One stage now writes the backlog and the plan you approve, so a run makes one fewer model call.
Runs say when web search is off and how to connect Docker for it, at the start and at the end.
Change a saved provider's model, label, or base URL right in Settings — existing projects pick it up on their next run, no need to recreate anything.
A failed or paused stage now shows one clear card — pick the provider, add an optional note, and go — instead of a hidden choice split across two different-looking controls.
A message you send mid-run no longer disappears from view — it shows right away as waiting, not only once the agent's next turn finishes.
Research, Vision, Architect, Sprint Planning, and Finalization now always produce and register their document — a stage can no longer finish having only talked about it.
Projects is now a live board — Needs you / Working / Done — instead of a flat list, so it's always obvious what needs your attention.
Edit, reorder, remove, or add backlog tasks right at the approval screen — no need to regenerate the whole plan over one wrong task.
Get notified the moment a plan needs approval, a run fails, or a build finishes — leave Atlas Studio running in the background.
A real progress bar during a run — tasks done, or planning stages reached — instead of just a spinner for however long it takes.
The finished-run summary now suggests concrete next steps (auth, edge-case tests, production config) — click one to draft a follow-up prompt.
A real Appearance setting in Settings — no longer limited to whatever your OS happens to be set to.
Pick a cheaper or local model for the narrative planning stages and a stronger one for the actual implementation, per project.
Every agent now has the same icon and color everywhere it appears — the run view, project history, and the board — not just the Agents tab.
Retuned contrast in both themes so a run that needs you actually stands out, instead of a finished one visually shouting louder.
An agent no longer claims something builds or works unless it actually watched that happen this run — it says plainly when something is unverified or broken.
The Architect now writes an AGENTS.md file at the project root — stack, conventions, and exact commands — that other AI coding tools can read too.
Every clickable card and option now has a visible focus ring and works from the keyboard, not just the mouse.
Asking a follow-up question after a finished build used to hide the terminal and cost breakdown for good, even after the follow-up finished.
The sidebar's "Trial x/y" is now clickable — jumps straight to a full FAQ entry on what's included and how to install on Mac/Windows.
Public downloads are trial-limited to one project per provider type (Anthropic, OpenAI, Local) — full pipeline on each, nothing held back.
API keys are now encrypted with the OS's built-in secure storage — fixed a startup crash on some Mac installs.
A failed stage now leads with plain-language guidance ("check your API key", "provider server hiccup") — raw details one click away.
A provider server hiccup retries automatically before failing the stage, instead of stopping the whole run.
The full project build (Maven, npm, Gradle, ...) must pass before a run finalizes — failures go back to the developer.
The completion summary lists the exact commands to build, run, and test your project.
Costs use current internet-checked model prices; local models are always $0.
A true PTY shell — cd sticks, colors and interactive programs work.
Pause mid-run, swap models, resume the exact conversation — no retyping.
Tokens per agent and estimated cost when a build finishes ($0 for local).
Unchanged file reads deduped, smarter truncation, deeper prompt caching.
Forced tool calls, repetition detection, and stricter compact prompts.
Read the vision/architecture docs while agents work — steer informed.
Research ✓ Vision ✓ Developer ⟳ — the whole pipeline at a glance.
Projects page shows totals and cost; the sidebar tracks the live backlog.
Give any agent standing instructions that apply to every future run.
Nothing gets built until you review and approve the backlog.
Every completed task is committed automatically.
Push new projects, and push follow-up commits later.
Two install-time warnings you'll likely see on a fresh download — both expected, both harmless.
This is macOS Gatekeeper, not a bad download — it flags any app downloaded outside the Mac App Store that isn’t notarized yet. Fix: open Terminal and run xattr -cr /Applications/Atlas\ Studio.app (after dragging Atlas Studio into Applications), then open it normally.
This is Windows SmartScreen flagging an installer without a paid code-signing certificate — not a virus warning. Fix: click "More info", then "Run anyway" to continue installing.
Every report becomes a public issue on GitHub, so you can follow it there.
The same FAQ lives inside the app too, one click from the sidebar.
As many as you need. Each terminal tab is its own shell in the project folder, with your normal login setup. Open more with "+ New terminal" and close a tab with its ×.
Web search runs through a small local search engine (SearXNG) that needs Docker. Without it, agents rely on memory for library versions and best practices, which can be out of date. Connect Docker and start the search engine for better results; the notice goes away once it is running.
It takes the tokens each provider reported for the run and prices them with published list prices, showing the low and high end when two price sources disagree. With web search on, the sources it found are listed under the figure. Cached input can lower the real bill, and it is an estimate, not an invoice.
Yes, once no stage is running. Open the project in IDE mode to create, rename, copy, paste, upload, and delete files and folders, and to edit and save text files. While a stage runs, files are read-only; pause the stage to edit.
Choose "Import an existing folder" on the New Project page and pick the folder. Atlas Studio maps the code and documents its architecture in atlas/ and AGENTS.md without changing the code. Add a note to start a build straight away, or leave it empty to read the documentation first.
Only at the end of a build, after the build and tests pass. It starts the app, checks that it answers, and stops it again. Agents are not allowed to start servers while they work, so a stuck server cannot block a run.
Each run writes a dated report to atlas/run-reports in the project folder: tokens per stage, time, backlog status, problems, and follow-ups.
It depends on the project. A small project gets a flowchart; a larger one also gets a high-level design, a low-level design, and a sequence diagram, written in Mermaid so they render in the app and in the IDE preview.
A desktop app that runs a small team of specialist AI agents — Researcher, Architect, Senior Developer, Tester, and more — through a real software development lifecycle from one prompt, and pauses for your approval before anything gets built.
3 trial projects, any provider mix — all 3 with Anthropic, all 3 local, or any combination, no restriction on which provider each one uses. Every agent and every pipeline stage runs at full strength on each: real research, real planning, real git commits. Once all 3 are used, buy the full version (one-time, price shown above) to keep building — it's a separate, uncapped download.
Click Buy above and pay through Razorpay's checkout — card, UPI, or netbanking in India; card or PayPal outside India — and the same page shows you a download link plus a permanent license key right after. The download link is only active for 10 minutes, but the key never expires — save it, and use it on atlasstudioai.com/redeem anytime you need a fresh download link (a new machine, a reinstall, whatever).
Go to atlasstudioai.com/redeem and paste the license key you got at checkout. It'll give you a fresh 10-minute download link for the exact platform you bought (Mac Apple Silicon, Mac Intel, or Windows — a key only unlocks the one you paid for). No account or sign-in needed, just the key.
Go to atlasstudioai.com/find-license and enter the Payment ID from your Razorpay confirmation SMS/email (or your bank/UPI app's transaction reference) — it looks the key up instantly and gives you a fresh download link too. If you don't have that either, email writetous@atlasstudioai.com with a screenshot of the payment and we'll look it up manually.
Yes — Atlas Studio itself doesn't charge for the trial, but you connect your own model provider. Cloud models (Anthropic, OpenAI) bill you directly for tokens used, shown as a live-checked cost estimate. Local models (Ollama, LM Studio, ...) have no per-token cost at all.
Only to reach whichever model provider you've connected. With a fully local model server, Atlas Studio can build entirely offline.
Yes — everything runs on your own machine. Your project files never leave your computer except to whichever model API you explicitly configure, and API keys are encrypted with your OS's built-in secure storage, never in plain text.
So nothing gets built without your sign-off. Atlas Studio researches, writes a vision, picks an architecture, and proposes a task backlog — then stops. You can edit or regenerate the plan before a single line of code is written.
Finalization always runs the project's real build/compile command before a run can complete. If it fails, the exact error goes back to the Senior Developer automatically to fix and re-verify — it doesn't finish a run with a broken build.
macOS (Apple Silicon and Intel) and Windows, both downloadable above. Linux is on the roadmap.
No — this is macOS Gatekeeper, not a bad download. It shows that message for apps downloaded from outside the Mac App Store that aren't notarized by Apple yet, not just for the first-time-open warning you might expect. It's a one-time fix: open Terminal and run xattr -cr /Applications/Atlas\ Studio.app (after dragging Atlas Studio into Applications), then open it normally. We're working on proper notarization so this step goes away.
No — this is Windows SmartScreen flagging an installer it doesn't yet recognize, which happens to any app not signed with a paid code-signing certificate. Click "More info", then "Run anyway" to continue installing. We're working on getting a proper signing certificate so this step goes away.
Atlas Studio is versioned with git tags — check this page any time for the latest release, or watch the app's own Release Notes tab after installing.
Research, plan, build, test, ship — with you approving every step of the way. Buy the full version, or try it first with 3 free trial projects.