Sparse

Every task, routed to the right model.

Claude, GPT and Gemini are each best at different things, but most coding agents stick to one. Sparse plans your task and sends every part to the model that fits it.

Download for macOS

Free. macOS on Apple Silicon. Updates itself.

Download for Windows

Free. Windows on x64. Updates itself.

npm install -g sparse-cli
All downloads, the CLI and the MCP server

One assistant, every model

One OpenRouter key gives Sparse the whole catalog: Anthropic, OpenAI, Google, DeepSeek, Moonshot, Z.ai and more. Prefer local? Point it at Ollama.

You review every change

Edits come back as a diff you accept or undo, file by file. Work on a copy keeps them in a git worktree until you do.

See every decision

Each model pick is logged with its reason, and routing improves from the outcomes of your own runs.

Chat or code

Questions, emails and research get one well-routed answer. Coding tasks get a full pipeline: plan, edit, validate, review.

Many models. One plan.

Sparse isn't another model. It's the planner and orchestrator that decides which model does what.

›Add team billing with StripePlannersplits the work, judges difficultyBilling data modelarchitectureSettings page UIroutine UIWebhook testsroutine testsOrchestratorroutes, runs side by side, checksAnthropic (Claude)OpenAI (GPT)Google (Gemini)DeepSeek (DeepSeek)Moonshot (Kimi)Z.ai (GLM)✓ Tests pass · one diff to review
  1. 1 / 3

    A planner that thinks first

    Sparse reads your whole request before writing a line. It breaks the work into steps, maps what depends on what, and works out how hard each one really is.

  2. 2 / 3

    An orchestrator that picks the model

    Each step goes to the model best suited to it, from Anthropic, OpenAI, Google and more. The data model gets a frontier model. The settings page and the tests get fast, cheap ones. Steps run side by side.

  3. 3 / 3

    One change, checked

    The pieces come back together as one change, checked against your tests before you see it. If a step falls short, a stronger model takes it over. You never have to pick a model.

Under the hood

What Sparse is for, how it decides, and what runs where.

The objective

The quality of the best model on every change, paying for it only where it matters.

Most of a coding session isn't hard. Renames, boilerplate and "where is X?" don't need a frontier model. Architecture, tricky bugs and changes that ripple across a codebase do.

Running everything on the most expensive model wastes money. Running everything on a cheap one wastes your time with bad diffs. Sparse makes that call for every step, so you never have to.

The approach

Conversation or task

Every prompt is sorted first. A question gets one well-routed answer with read-only and web tools. Work on your repo goes through the full pipeline: plan, edit, validate, review.

Judge the shape, not just the size

Each step is scored on the kind of thinking it needs: deep reasoning, long context, taste across many files, adversarial checking or open exploration. Two steps of the same size can need very different models.

Quality floors by blast radius

Anything that touches architecture, or that several other steps depend on, gets a top-tier model whatever it costs. Saving money never comes from the parts that can break everything.

Escalate instead of guessing

Every step is checked by validators. One that fails moves up to a stronger model, and a strong-model integration pass reviews the merged result before you see it.

Learn from your repo

Outcomes feed back into routing. Routes that work on your code get preferred, routes that keep failing get avoided, and both fade over time so old results never lock it in.

Your call when you want it

Pin a Lead for planning and hard work and a Workhorse for the routine, or let Sparse balance cost and quality on its own. Every pick is logged with its reason.

The architecture

One engine on your machine, many ways in. Every client talks to the same local daemon, so a run you start in the terminal shows up in the app.

  1. Where you work

    • Desktop app
    • CLI
    • VS Code and Cursor
    • Any MCP host
  2. Local daemon

    Runs on your machine and keeps long runs and loops going after you close the window. Enforces your monthly spending limit.

  3. The engine

    • Router

      conversation or task, shape, difficulty

    • Planner

      sub-tasks, dependencies, quality floors

    • Orchestrator

      parallel runs in isolated git worktrees

    • Validators

      tests and checks, escalation, integration pass

    Edits come back as a diff you accept or undo. With Work on a copy, they stay in a git worktree until you do.

  4. Models

    and the rest of the catalog. Your code goes only to the model doing the step.

Stop choosing models. Just build.

Download for macOS

Free. macOS on Apple Silicon. Updates itself.

Download for Windows

Free. Windows on x64. Updates itself.

Want the details first? Read the docs.

About

Sparse is built by Efe Acar. It started from a simple observation: most of a coding session isn't hard, so it shouldn't all run on the most expensive model.

Get in touch
efeacar2705@gmail.com
GitHub
EfeAcar6431
Found a bug or want a feature?
Open an issue on GitHub