GOAT$10/mo with $70 credits~50 models(best low cost coding plan)

Meet the coding agent
built for open models.
Command Code.

Downloadnpm i -g command-code

cmd.mp4

Start with $1/mo or $10/mo, get $70 free credits.

GOAT$10/mo with $70 credits~50 models

Explore all features

Docs

Always learning

Every accept, reject, and edit is a signal auto-generates into project-level skills.

A mode for every moment

Interactive CLI. Headless `-p`, `--yolo`. Background sandbox.

Best in class tools. Built in

Read and shell tools that lead the token efficiency frontier. File ops, grep, extended thinking. The full stack, in your hands.

It remembers. So you don't have to

Custom /agents and persistent /memory carried across every session.

Hackable. Out of the box

Reusable /skills, /commands, /mcp servers, plugins. Built to be hacked.

Even better, together

/share sessions. npx taste push/pull. Taste, across teams.

Your design partner

One command, seventeen modes. Audits, recolors, and reships your interface.

~/project

01Your coding taste `npx taste`02Interactive, headless, sandbox agents 03Skills, reviews, bundled agentic tools 04Context engineering with /agents, /memory 05Hackable /skills, /mcp, /commands, and plugins 06Collaborate /share sessions, npx taste push/pull 07Your design partner `/design`Explore

//Always learning. Every accept, reject, and edit is a signal auto-generates into project-level skills.

~/project

//Take a closer look.

[]Play01Command Code. In 74 seconds

[]Play02Hooks in Command Code block unsafe commands and enforce project rules

[]Play03Introducing Cloud sessions with sandboxes in Command Code

[]Play04Working with ACP (agent client protocol) in Command Code

[]Play05Use /add-dir to work across multiple directories in Command Code

[]Play06Planning and Implementation models in Command Code

[]Play07Remote Control now available in Command Code

[]Play08/loop is now live in Command Code

[]Play09You can now use Shift+Tab to bypass permissions

[]Play10Command Code supports Git worktrees

[]Play11Stop copy-pasting errors. IDE diagnostics are quite strong

[]Play12Mods are more powerful than plugins

[]Play13Fork your existing sessions and create new ones

[]Play14Connect to MCP servers to query data, use tools and automate workflows

[]Play15Open an external editor for your multi-line and longer prompts

[]Play16Plan review mode is best to comment, revise, and approve your plans

[]Play17Resume and pick up right where you left off. /resume command

[]Play18Paste images as your chat input in Command Code

[]Play19Make screenshots your prompts. No more describing just attach them

[]Play20Try the new a11y mode with /design in Command Code

[]Play21Writing tests have never been this easy

[]Play22Pull requests are easier now

[]Play23Are code reviews blocking? No need to get slow!

[]Play24Think before you actually ship: Plan Mode!

[]Play25Stop repeating the same prompt in every coding session. Do this instead!!

[]Play26Does your code have your coding taste?

[]Play27Save your favorite prompts forever and reuse them anytime!!

[]Play28Do this to bring back any message or code instantly

[]Play29Build your custom agents for your work

[]Play30Does your AI has permission modes?

[]Play31Your AI agents will now never forget their sessions and context

[]Play32Share your AI models usage with others. Tokenmaxing, cost, performance

[]Play33Refactoring with Command Code hits different

[]Play34Share your coding session with others

[]Play35Getting file context for your AI agent

[]Play36Understanding your codebase. (Next.js +TypeScript storefront)

[]Play37Tool repairs and fixing tool calling issues in Command Code

[]Play38/goal in Command Code

38 demo videos. More on YouTube

command code desktop app

Code with your agent. Review changes, run commands, and preview your app in one workspace.

Download for macOS

Available on macOS, Linux & WindowsRead docs

Stable first paintdesktop

SearchK

New thread

Pull requests

desktopStable first paint

landing

docs

Command Code

Desktop preview

trace the startup flash, fix it, and prove the first paint is stable

Worked 11

Shellpnpm test src/layout --run

Searchtheme initialization in src/renderer

Read filesrc/renderer/layout/Sidebar.tsx

Read filesrc/renderer/theme/useResolvedTheme.ts

Browserlocalhost:3000 startup frame

EditedSidebar.tsx+8-2

EditedSidebar.test.tsx+31-0

Agents3 background checks complete

Background taskRendererTracing the first paint

Background taskTheme auditChecking persisted tokens

Background taskTest runnerVerifying cold starts

Root cause found in the first render. The sidebar painted before

the persisted theme resolved, then swapped tokens one frame later.

I gated the shell on theme readiness and added a regression test.

Result: zero color flash, zero layout shift, 138 ms cold start.

Write a message...

desktopmainAsk for approvalV4 Flash (latest)

Terminal 1+Setup Run

Sidebar.test.tsx (18 tests) 94ms

Test Files 1 passed (1)

Tests 18 passed (18)

Duration 1.21s

ChangesBrowserTerminalFilesdesktop

Commit message...Commit

src/renderer/layout/Sidebar.tsxM

3939 export function Sidebar({children}) {

4040 const persisted = usePersistedTheme();

4141 const theme = useResolvedTheme();

4242 const ready = theme !== undefined;

42-return <Shell tone={theme} />;

42+if (!theme) return null;

43+return <Shell tone={theme} />;

44+data-ready={ready}

4444 return children;

4545 }

4646 export default Sidebar;

4747

4848 type SidebarProps = LayoutProps;

code without taste is slop

//Slop. Generic. Not yours. LLMs generate sloppy code. Not your patterns. Not your taste. What if you could code like you think. Flow state accelerated with your coding taste.

harness, not the model

//Open models work best in Command Code. Free 1M tool call repairs per 1T tokens. Best in class 99%+ cache hit rates. The best harness, leading the Pareto frontier on the TEF token efficiency benchmarks.

[ Input ]

[ Thinking ]

Continuous learning.every accept, reject, and edit, a signal.

Auto-generating skills.patterns become reusable skills, no rules to write

Push to team.one command, the whole team pulls instantly

[ Output ]

npm i -g command-code

Docs

Without Taste

Other coding agents. The same play.

[ PROMPT ]> Build a cli to tell date

Building

[ WRONG ]Interrupted

Lhey use typescript

Blabbering Adding tsc

[ WRONG ]Interrupted

Lno, use tsup

Stackflowing Adding mocha

[ Wrong ]Interrupted

Li prefer vitest

Done!

[ f&5k ]> f&5k, use lowercased -v to for version,

Updated, now -v for version.

You can run the app using `npm run dev`

[ s#!t ]> s#!t, i always prefer pnpm

Stuck. Sloppy AI.> leave it, i'll do it myself!

> learn something from me for a change

Mmmmm. Taste.

Command. After it learns you.

> Build a cli to tell date[ PROMPT ]

Building cli, let me check your taste

[ npx taste pull ]Taste

Learned from youL

Using your taste, I see you prefer:

TypeScript for CLI

Commander and tsup

Vitest for tests

You prefer pnpm but do `npm link`

You like lowercased `-v` for cli version

TODO

Using taste, learning, building...

[ Builds with your taste ]

Done!

Built a date cli, with TypeScript, tsup, vitest

Also linked using `npm link`

Run `date-cli` to try it out.

> oh wow, awesome! just what I wanted.[ continuous learning ]

> i just made an api route, can add /health routeLearned how you build APIs

Subscribe.
Be a GOAT.

[]GLM-5.2
[]GPT-5.6 Sol
[]Tencent Hy3
[]Qwen 3.8 27B
[]DeepSeek V4 Flash
[]DeepSeek V4.1 Flash
[]Kimi K2.7 Code
[]MiniMax M3
[]Gemini 3.7 Flash
50+ models

[]

GOAT

$10/mo

Best value plan
on the market.

[]Z.ai
[]OpenAI
[]Tencent
[]Alibaba
[]DeepSeek
[]Moonshot AI
[]MiniMax
[]Google

50+ models

Subscribe to GOAT

Built to scale.

See full pricing

Go$1/mo Provider$15/mo Pro$20/mo Max 10$100/mo Max 20$200/mo

//Community

Developers love it. Founders, too.

Follow on X

Follow on LinkedIn

Learns your taste

Command Code learns my taste. After a week, it stopped making the mistakes I kept fixing in other agents. The diffs feel like a senior engineer who already read the codebase.

[Zeno Rocha]

Zeno Rocha

Founder & CEO Resend

Open models, premium output

Command Code is the first agent where I trust open models in production. The harness is so solid I had to double check I was still on DeepSeek Flash. Shipped multiple CLIs and a full Redwood app for $2.

[David Thyresson]

David Thyresson

GP PWV RedwoodSDK contributor

[]

[]

[]

[]

Start building.With your taste.

Get started

Install. Sign in. Code.

//faq

Questions, answered.

Everything that usually comes up before a team installs. Still curious? Read the docs, or join the Discord.

How is Command Code different from Cursor or Copilot?

Command Code is the best coding agent for open models. It lives in your terminal and continuously learns your coding taste. Powered by taste-1, it ships, fixes, tests, and refactors with the patterns you keep, and forgets the ones you delete.

Why is Command Code the best coding agent for open models?

Because we made them work. "Open model bad at tool calling" is a harness problem, so Command Code validates and repairs every tool call in flight, about 1M repairs per 1T tokens, free on every plan. Our read and shell tools lead the token efficiency frontier benchmarks across ten harnesses, and our prompt caching holds 99%+ cache hit rates, the best in the industry. DeepSeek, Kimi, GLM, Qwen, and MiniMax all run on it, proven at 30T scale, 130T+ tokens served, by 100K+ developers and 40K+ paid customers.

What does "learns my taste" actually mean?

Every accept, reject, and edit is a signal. Command Code distills those into project-level /skills and personal /memory, so the next session opens with the conventions you already prefer. No rules to write, no prompts to maintain.

Which models can I use? Can I bring my own?

Every model listed in our docs ships out of the box Anthropic, OpenAI, Google, xAI, DeepSeek, Qwen, Kimi, GLM, MiniMax, and more. New vendors land regularly.

Is my code used for training?

No. Never. Your code, your skills, and your memory stay on your machine. Command Code never trains on your work. See the Privacy Policy.

How do teams share taste?

npx taste push to publish a project skill, npx taste pull to install one. Skills are open files in your repo review them in PRs like any other code.

What does it cost?

Free tier for solo developers. Pro and Team plans add seats, more compute, and shared taste registries. See Pricing for the full breakdown.

//ship