desktop · macOS · Windows · Linux

A universal agent for tasks of any complexity

A platform for LLMs: its own terminal, computer use on screen and a built-in browser. On the very same model, maracuya solves twice as many Terminal-Bench 4.0 tasks as OpenCode and four times as many as Cursor.

Download for macOSChoose a desktop version

macOS 13+ · Apple Siliconan app for macOS, Windows and Linux·other systems

maracuya chat: the agent works out why the build fails

app mockup — you can click and scrollinteractive mockup — on desktop

Examples

What maracuya can build

Landing page in zoomer style

One prompt — “redesign maracuya.ru in zoomer style”. One pass produced this landing. No follow-up edits. Total cost — $1.47

Landing page in Soviet poster style

Three prompts: one pass for the site, one for small fixes, one to rewrite copy in Mayakovsky’s style. Total cost — $2.23

Game “Divine Strategy”

Prompt 1: “I need a game with no backend, progress in localStorage. A 3D browser strategy: economy, building, divine powers, opponents, 10+ hours of gameplay

Build a three-dimensional site with motion design

Prompt: “For an example of how well our solution, Maracuya AI, performs, I need one more website. It has to be motion design — a three-dimensional site, in Russian and English, which I will embed into our current landing (see the new_landing repository).

Benchmarks

The two hardest tests for agents — and we're among the world leaders

We ran the terminal benchmark and the computer-use benchmark.

Terminal-Bench 4.0 — engineering and scientific tasks in code

one attempt per task

Only OpenAI's and Anthropic's flagship models paired with Claude Code and Codex are ahead. We ran the benchmark on our own rig; its validity is confirmed by a Claude Code + Claude Opus 5 run — 56% against 55% at Artificial Analysis.

maracuyaour runs of other platformsArtificial Analysisfully solved, % · 21 Sep 2026

Claude Code + Claude Fable 5.1 (max)58

Codex + GPT-6 Astra (max)56

Claude Code + Claude Opus 5 (max)55

Codex + GPT-5.6 Sol (max)37

maracuya + DeepSeek V4.1 Flash34

Muse Code + Muse Spark 1.3 (max)32

Kimi Code CLI + Kimi K321

Grok Build + Grok 4.6 (xhigh)18

Claude Code + Qwen3.8 Max17

OpenCode + DeepSeek V4.1 Flash17

Antigravity SDK + Gemini 3.8 Flash15

Codex + DeepSeek V4 Flash 073111

Codex + DeepSeek V4 Pro 081310

Cursor + DeepSeek V4.1 Flash8

Our own comparison: identical environment, one model — DeepSeek V4.1 Flash, reasoning effort high.

platform

fully solved

checks passed

maracuya

34%

87%

OpenCode

17%

76%

Cursor

70%

87% of checks passed means one simple thing: every task was solved almost completely.

OSWorld 2.0 — the agent works on the computer screen

official container

General tasks where the agent drives applications and interfaces rather than writing code. We fully solved 20.2% of the tasks and almost half more partially, for an average score of 47.8%. On the same DeepSeek V4.1 Flash model, the benchmark's stock platform solves 3.7%.

maracuyasame model without maracuyaofficial leaderboardfully solved, %

Claude Opus 5 (max)31.4

GPT-5.6 Sol (max)27.3

Qwen-CUA-Max21.2

Claude Opus 4.8 (max)20.6

maracuya + DeepSeek V4.1 Flash20.2

Qwen-CUA18.5

Claude Opus 4.7 (max)18.2

GPT-5.5 (xhigh)13.0

Claude Sonnet 4.6 (max)8.3

LongHorizon + Qwen 3.7-Plus8.3

Kimi 2.64.6

MiniMax M34.6

DeepSeek V4.1 Flash without maracuya3.7

Qwen 3.7-Plus2.8

Drives any applicationThe agent sees the screen, moves the cursor and presses keys — it works where there is neither an API nor a command line.

Built-in browserOpens links, fills in forms and takes page snapshots in the same window. Where a login is needed, sign in yourself — the agent carries on already logged in.

A rare ability for platformsCoding platforms usually don't take this benchmark: OpenCode and Cursor have never entered it on their own.

Economics

One benchmark run on a flagship model costs about $2,000. Ours — $13.70

The benchmarks show it: right now we trail only Claude Code and Codex — and only slightly. But both of those platforms enter the test paired with flagship models, and a single run like that costs roughly one and a half to two thousand dollars.

We ran Terminal-Bench 4.0 for $13.70 and OSWorld 2.0 for $16.50. The price gap isn't a few times over, it's hundreds of times: the result rests on the platform, not on the most expensive model.

What a platform is

Artificial intelligence has two layers. The second is less visible, yet it is what turns a model into a tool

The first layer is the models themselves — everyone knows about them. The second is everything around the model: the agent loop, tools, skills, sandboxes for running code, chat, memory. There is no established term; we call this layer the platform. Examples: Claude Code, Codex, Cursor.

They are usually called coding platforms, though in practice they are general-purpose: as a rule, the platform solves a task by writing and running code.

What's inside

What the platform in the app is made of

Your files, not copy-paste

Open a project folder — the agent reads it whole, edits files, creates new ones and shows a diff before saving.

A terminal in the same window

Builds, tests, migrations, git. The agent runs commands and reads the output itself — you see every line. This is the same terminal we ran Terminal-Bench 4.0 on.

Several agents at once

Hand a long task off to the background and run a second one in parallel — each has its own work plan and its own branch.

Skills and connectors

Plug in your MCP servers, databases and services. A skill is a set of rules and tools for a specific job.

Computer use

The agent drives applications and browsers the way a person does: it sees the screen, clicks, types. The built-in browser is in the same window.

Your key, your models

A maracuya subscription or your own API key from the provider. Token usage and cost are visible in the app.

Models

You choose the model — in one click, mid-conversation

You can work on the models we offer and pay us for tokens. Or enter your own API key and pay the provider directly.

DeepSeek

OpenAI

Google Gemini

Mistral

xAI

Moonshot

Z.ai

your own key

Permissions

Three trust modes. Switch at any moment

Ask for permission

Always ask before editing files outside the project or going online. The default mode.

Decide for me

Ask only about risky actions. The agent handles routine on its own and reports to the chat.

Full access

No limits: any files on the computer and the whole internet. For trusted projects.

You can see what the agent is about to do

Reasoning, the list of actions, the work plan and a diff of every file — all in the conversation. Adjust the plan on the fly or stop the run with one button.