GenGrades
EN中文

Command Code: $1 coding agent, GOAT plan and taste learning

Published Updated

Command Code logocommandcode.aiSubscriptionCLIDesktopPluginAPI
The Command Code homepage: a coding agent built for open models, installed with npm or as a desktop app.
The Command Code homepage: a coding agent built for open models, installed with npm or as a desktop app.Screenshot of commandcode.aitaken

What Command Code is

In one or two sentences, before anything else.

Command Code is a terminal coding agent from Langbase, Inc., trading as Command Code and led by Ahmad Awais, that says it learns your coding "taste" from the changes you accept, reject and edit. Announced with 5 million USD in seed funding on February 25, 2026, it now focuses on low-cost open models, with plans from 1 USD a month.

Main features

What it can do, one sentence each, and whether we actually tested it.

  • Taste learning (taste-1)

    Turns each accept, reject and edit into a signal; a model called taste-1 sorts the signals into preferences, and an LLM writes them as Markdown under .commandcode/taste/ for later turns to read.

    Per Command Code's taste docs. Command Code says taste-1 is a classifier, not a text model; its claimed drop in correction loops is its own figure.

  • Shareable taste packages

    npx taste push publishes a project's taste to commandcode.ai and npx taste pull installs someone else's, so a team can share its conventions.

    Per Command Code's launch post and docs. Packages can be public or private.

  • Open and closed models in one picker

    Lists DeepSeek, GLM, Kimi, Qwen, MiniMax, MiMo and other open models next to Claude, GPT, Gemini and Grok; the default is DeepSeek V4 Flash.

    Per Command Code's model reference, checked October 17, 2026. Its pricing FAQ says premium models such as Claude join from the 20 USD Pro plan.

  • Tool-call repairs for open models

    Fixes malformed tool calls from open models inside the agent loop at no extra charge.

    Command Code claims about one million repairs per trillion tokens and 99%+ cache hit rates; we found no independent measurement.

  • Bring your own provider (beta)

    Connects 150+ providers, local servers such as Ollama or vLLM, and ChatGPT Plus or Pro and GitHub Copilot subscriptions; these requests go from your machine straight to the provider.

    Per Command Code's BYOK docs. BYOK arrived in beta on August 21, 2026.

  • Desktop app and IDE extension (beta)

    A desktop app for macOS, Windows and Linux with plans, Git changes, a terminal and browser preview; an extension for VS Code and Cursor; ACP for editors such as Zed.

    The desktop docs warn of bugs and say the Windows installer may be unsigned.

  • Five permission modes

    From asking before every file write or shell command to a yolo mode with no prompts; headless runs (cmd -p) block writes unless you pass --yolo.

    Per Command Code's security docs. Deny rules apply in every mode.

  • Plan mode, worktrees, /loop and mods

    Review a plan before edits, run parallel sessions in Git worktrees, repeat a task on a schedule with /loop, and extend the terminal interface with mods.

    Per Command Code's blog and changelog, September and October 2026.

Who it is for

The people it helps and the problems it solves, and where it ranks.

Built for

  • Developers who want a terminal coding agent on cheap open models such as DeepSeek, GLM, Kimi or Qwen
  • Developers who keep correcting the same style and tooling choices in agent output
  • Small teams that want to share coding conventions as files rather than a long rules document
  • Developers who want to try a paid coding agent for 1 USD a month before committing to a 20 USD plan

Problems it solves

  • Fixing the same naming, tooling and structure choices every time an agent writes code
  • Rules files such as CLAUDE.md or AGENTS.md going stale as the codebase changes
  • Paying 20 USD or more a month before knowing whether a coding agent suits your work
  • Open models making malformed tool calls in long agent loops

Graded in

What it is good at

Public tests other people ran: what they gave it and how it did.

  1. Case 1

    In AICodeKing's September 24, 2026 YouTube test: the desktop app on DeepSeek V4 Flash, on the 1 USD Go plan

    Input

    In a small throwaway repository, add search and status filters to a task list: case-insensitive search that works with the status filter, an empty state, and a stated test plan.

    What happened

    Plan review caught a reversed text comparison in the first plan before any code was written. The agent worked around a missing Node path by itself, added a forgotten "all statuses" option, and all 11 tests passed; some alpha features had rough edges. One run, per the video description and OpenClaw Database's summary; the description links to Command Code with a campaign tag.

  2. Case 2

    In Command Code's own launch demo (February 25, 2026): the same prompt in Claude Code and in Command Code

    Input

    Build a cli to tell date (both agents had seen Ahmad Awais build command-line tools before).

    What happened

    Command Code used TypeScript, Commander.js, version 0.0.1 with a lowercase -v flag and ISO dates, choices it had learned from earlier sessions; Claude Code printed the date from a plain script. This is the vendor's own demonstration, not an independent test.

  3. Case 3

    Per Md Zia Uddin's notes on his own setup (using it since March 2026): Command Code as a terminal agent next to Cursor

    Input

    His example prompt: add a team member page matching github-copilot.md. He also runs headless cmd -p jobs and /learn-taste after real edits.

    What happened

    He says the learned taste compounds from his accept and reject signals and values shareable taste packages. It learns the wrong habits when he approves bad diffs, the cmd command clashes with the Windows shell (he uses cmdc), and Cursor stays his daily editor agent.

Where it falls short

What it does badly, with the evidence, and what we could not settle.

Weak

Weak

  • Almost no independent testing.

    Evidence: We found one hands-on video (AICodeKing, September 24, 2026) and one user's notes, and no benchmark against Claude Code, Codex, OpenCode or Cursor. Headline numbers such as code 10x faster, 5x fewer bugs and correction loops falling from 4.2 to 0.4 edits are Command Code's own, with no published method.

  • Not open source, despite a promise.

    Evidence: Command Code's GOAT plan post (August 5, 2026) said it would open-source Command Code later that month. On October 17, 2026 its public GitHub repository held a readme, mods and issue templates, with no licence and no agent source code.

  • Prompts and code snippets are kept for up to 30 days by default.

    Evidence: Its privacy policy (updated October 7, 2026) retains prompts, attached files, code snippets and outputs for up to 30 days unless you run with CMD_ZDR=1, and says free or promotional models may run under the provider's own terms. Command Code says it does not train on your source code.

  • Beta apps and telemetry on by default.

    Evidence: The desktop app and the VS Code and Cursor extension are in beta, and the desktop docs say the Windows installer may be unsigned. The CLI sends anonymous telemetry unless you set DO_NOT_TRACK=1 or telemetry: false (telemetry docs).

  • Not for code that must not leave your machine, or for anyone who needs an open-source agent today.

    Evidence: On a Command Code plan, requests go to Anthropic, OpenAI or, through gateways such as Vercel, Cloudflare and OpenRouter, to model hosts in the US, EU or Singapore (privacy policy and pricing FAQ). Such teams should wait, or use the MIT-licensed OpenCode with a local model.

Uncertain

Uncertain

  • Whether there is a free tier.

    Why unsure: The homepage FAQ mentions a free tier for solo developers and the launch post offers 10 USD in sign-up credits, but the pricing page lists no free plan: the cheapest is Go at 1 USD a month plus a processing fee, and a few preview models cost no credits.

  • How many people use it.

    Why unsure: The homepage claims 100K+ developers and 40K+ paid customers; a post by Ahmad Awais quoted on the pricing page says 53K active paying customers. Neither figure is independently confirmed.

  • What the GOAT plan includes.

    Why unsure: On October 17, 2026 the plan card said 80 USD in credits, the comparison table 70 USD, and the August 5 launch post 70 USD. Per-model allowances and promotions change every few days, per Command Code's blog.

  • Where taste learning runs.

    Why unsure: The pricing FAQ says taste data stays on your machine only, the privacy policy says taste profiles sync to your account if you opt in, and the taste docs describe taste-1 as a model Command Code pays to run.

  • Which date counts as the launch.

    Why unsure: The changelog starts with v0.1.0 on November 26, 2025, the launch post is dated February 25, 2026, and Ahmad Awais writes "We launched in May", probably meaning the open-model plans.

Price, limits and formats

The numbers, each with its source and the date we checked it.

SpecValueNote and source
MakerLangbase, Inc., doing business as Command Code (San Francisco)

Source: commandcode.aichecked

Founder and CEOAhmad Awais

Source: commandcode.aichecked

First CLI releasev0.1.0, November 26, 2025

Source: commandcode.aichecked

Public launch and fundingFebruary 25, 2026; 5 million USD seed round led by PWV (Tom Preston-Werner)

PWV's lead role is named on Command Code's launch announcement page (commandcode.ai/launch).

Source: commandcode.aichecked

Plans, monthly, plus a processing feeGo 1 USD (10 USD credits); GOAT 10 USD (80 USD credits on the plan card); Pro 20 USD; Max 100 USD (150 USD credits) or 200 USD (300 USD credits); Teams 40 USD; Provider API 15 USD

Source: commandcode.aichecked

CLI and editor extensionCLI installed with npm i -g command-code; VS Code and Cursor extension in beta

Source: commandcode.aichecked

Desktop appBeta since September 16, 2026, for macOS, Linux and Windows

Source: commandcode.aichecked

Training and retentionNo training on your source code; prompts, files and outputs kept up to 30 days unless CMD_ZDR=1

Source: commandcode.aichecked

TelemetryAnonymous usage telemetry on by default; off with DO_NOT_TRACK=1

Source: commandcode.aichecked

Source codeClosed; public repository holds readme, mods and issues, 4.1k stars on October 17, 2026

Source: github.comchecked

Common questions

Answer first, then why, an example, the limits and what to do next.

Who makes Command Code, and is it legit?

Command Code is made by Langbase, Inc., a San Francisco company that trades as Command Code and is led by founder and CEO Ahmad Awais. It announced a 5 million USD seed round led by PWV, the fund of GitHub co-founder Tom Preston-Werner, on February 25, 2026, and has shipped CLI releases almost daily since November 2025.

Limits

The funding round and user numbers come from Command Code itself; besides its own pages we found only a Signalbase funding listing dated March 4, 2026, and no news coverage.

How does Command Code learn my coding taste?

Command Code treats every accept, reject and edit as a signal and stores what it learns as Markdown files in your project. Per its docs, an LLM you choose turns your edits into structured signals, a model called taste-1 decides which are real preferences and how confident they are, and an LLM writes them to .commandcode/taste/. You can read, edit, push and pull those files.

Limits

One user reports that it learns bad habits if you approve bad diffs.

How much does Command Code cost?

Command Code starts at 1 USD a month for the Go plan, plus a processing fee, with 10 USD of model credits. GOAT costs 10 USD (80 USD in credits on the plan card), Pro costs 20 USD and adds premium models such as Claude, and Max costs 100 or 200 USD. Extra credits are sold at model cost and roll over.

Limits

The pricing page lists no free plan, although the homepage FAQ mentions one, and plan contents changed several times between April and October 2026.

Does Command Code train on my code?

Command Code says it does not train AI models on your source code. Its privacy policy (updated October 7, 2026) adds that prompts go to model providers, that it asks them not to train where they offer that option, and that prompts, files, code snippets and outputs are kept for up to 30 days unless you run with CMD_ZDR=1.

Limits

Free or promotional models may run under the model provider's own terms.

Is Command Code open source?

No, Command Code's agent was not open source on October 17, 2026. Its GitHub repository CommandCodeAI/command-code holds the readme, mods and issue tracker without a licence, although Command Code said on August 5, 2026 that it would open-source the agent later that month.

Next step

OpenCode is an MIT-licensed terminal agent with a similar 10 USD open-model plan.

Is Command Code better than Claude Code, Codex, OpenCode or Cursor?

No public test has run Command Code against Claude Code, Codex, OpenCode or Cursor on the same tasks, so nobody can say yet. Command Code is the cheapest way in at 1 USD a month and is built around open models; Claude Code needs Claude Pro at 20 USD a month; Codex comes with ChatGPT plans; OpenCode is free and open source with your own keys; Cursor is an editor with paid plans from 8 USD.

Next step

Command Code can also run GPT models through a ChatGPT Plus or Pro subscription via /connect (BYOK, beta).

More screenshots

  • The pricing page: Go at $1 a month with $10 in credits, GOAT at $10 with $80, Max from $100.
    The pricing page: Go at $1 a month with $10 in credits, GOAT at $10 with $80, Max from $100.Screenshot of commandcode.aitaken

What the public tests show

This page is built from what can be checked today: Command Code's site, docs, changelog, pricing and privacy pages, its GitHub repository, a funding listing, one hands-on video and one user's notes. Every price and date links to its source, and Command Code's claims are labelled as its own.

What it is and where it came from

Command Code is a coding agent you start with cmd in a terminal. Its privacy policy names the company as Langbase, Inc., doing business as Command Code; its about page lists Ahmad Awais as founder and CEO. It launched as an agent that learns your coding style, added open models in April 2026, and now calls itself "the best coding agent for open models".

DateEventSource
November 26, 2025CLI v0.1.0, first releasechangelog
February 25, 2026Launch post; 5 million USD seed round led by PWVblog, Signalbase
April 10 to 15, 2026Open models and new Go, Pro and Max planschangelog
July 22, 2026v1.0, a full rewritechangelog
August 5, 2026GOAT plan, 10 USD a monthblog
September 16, 2026Desktop app beta for macOS, Linux and Windowsblog
October 16, 2026CLI v1.84.0, the current releasechangelog

Its rise shows in public facts: 411 CLI releases in under eleven months, and 4.1k stars on its GitHub repository on October 17, 2026 (the repository holds the readme, mods and issues, not the agent). Its homepage claims 100K+ developers and 40K+ paid customers, which we could not check.

How taste learning works

Per the taste docs, every accept, reject and edit is a signal. An LLM you choose turns your edits into structured signals; taste-1, which Command Code describes as a classifier rather than a text model, decides which signals are real preferences and how confident they are; an LLM then writes them as Markdown under .commandcode/taste/, read back on later turns. You can edit those files and share them with npx taste push and npx taste pull.

The launch post claims correction loops for CLI scaffolding fall from 4.2 edits to 0.4 after a month. No method is published.

What it costs

PlanA month, plus a feeCredits
Go (open models)1 USD10 USD
GOAT10 USD80 USD (70 in the comparison table)
Pro (adds Claude and other premium models)20 USD80 USD
Max100 or 200 USD150 or 300 USD
Provider API15 USDPay as you go

Source: pricing page, October 17, 2026. Extra credits are sold at model cost and roll over. You can also bring your own keys or a ChatGPT Plus or Pro subscription through /connect (BYOK docs).

What public tests and users report

  • AICodeKing (YouTube, September 24, 2026) had the desktop app, on DeepSeek V4 Flash and the Go plan, add search and status filters to a task list. Per a summary, plan review caught a reversed comparison before coding and all 11 tests passed; some alpha features were rough. The video links to Command Code with a campaign tag.
  • Md Zia Uddin (his notes, using it since March 2026) runs it beside Cursor. He says taste compounds, but it learns wrong habits if he approves bad diffs, and cmd clashes with the Windows shell.
  • Command Code's launch demo gave Claude Code and Command Code the same prompt; Command Code used the founder's learned stack, Claude Code a plain script. That is the vendor's own demonstration.

No benchmark we found runs Command Code against other agents on the same tasks.

How it compares with Claude Code, Codex, OpenCode and Cursor

Command CodeClaude CodeCodexOpenCodeCursor
FormTerminal agent; desktop and IDE in betaTerminal and desktop appCLI, IDE, ChatGPT appsOpen-source terminal agentCode editor
Models50+ open and closedClaudeOpenAI75+ providers, local tooGrok on Plus; others on Premium
Cheapest paid1 USD a monthClaude Pro, 20 USDChatGPT Go, 8 USDOpenCode Go, 10 USD (free with own keys)Plus, 8 USD

Sources: Anthropic, OpenAI, OpenCode, OpenCode Go and Cursor.

OpenCode Go is the closest match on price. Command Code adds taste learning and tool-call repairs; OpenCode is open source, while Command Code's promised August 2026 release of its code had not happened by October 17.

Who should wait

  • Teams whose code must not leave their machines: plan requests go to model providers, some through gateways such as Vercel, Cloudflare and OpenRouter, and prompts and code snippets are kept up to 30 days unless you set CMD_ZDR=1 (privacy policy).
  • Anyone who needs an open-source agent today: OpenCode is MIT-licensed and runs local models.
  • Anyone who wants proof first: the speed claims are Command Code's own, and the only hands-on test we found is one video.

Sources

Sources

  1. Command Code homepage and FAQ (Command Code, checked October 17, 2026) (commandcode.ai), accessed
  2. Command Code plans and pricing FAQ (Command Code, checked October 17, 2026) (commandcode.ai), accessed
  3. Command Code privacy policy (Langbase, Inc. d/b/a Command Code, updated October 7, 2026) (commandcode.ai), accessed
  4. About Command Code: team and investors (Command Code, checked October 17, 2026) (commandcode.ai), accessed
  5. Command Code raises $5M: launch announcement (Command Code, checked October 17, 2026) (commandcode.ai), accessed
  6. Introducing Command Code: the coding agent that learns your taste (Command Code blog, February 25, 2026) (commandcode.ai), accessed
  7. Introducing the GOAT plan (Command Code blog, August 5, 2026) (commandcode.ai), accessed
  8. Introducing the Command Code Desktop App (Command Code blog, September 16, 2026) (commandcode.ai), accessed
  9. Use your ChatGPT subscription in Command Code (Command Code blog, September 29, 2026) (commandcode.ai), accessed
  10. Command Code changelog, v0.1.0 to v1.84.0 (November 26, 2025 to October 16, 2026) (commandcode.ai), accessed
  11. Command Code docs: Taste and taste-1 (checked October 17, 2026) (commandcode.ai), accessed
  12. Command Code docs: available models (checked October 17, 2026) (commandcode.ai), accessed
  13. Command Code docs: BYOK providers (checked October 17, 2026) (commandcode.ai), accessed
  14. Command Code docs: Desktop App (checked October 17, 2026) (commandcode.ai), accessed
  15. Command Code docs: IDE extension (checked October 17, 2026) (commandcode.ai), accessed
  16. Command Code docs: Security and privacy (checked October 17, 2026) (commandcode.ai), accessed
  17. Command Code docs: Telemetry (checked October 17, 2026) (commandcode.ai), accessed
  18. CommandCodeAI/command-code repository (GitHub, read October 17, 2026) (github.com), accessed
  19. Command Code raises 5.0M seed funding (Signalbase, March 4, 2026) (trysignalbase.com), accessed
  20. Command Code Desktop Agent App (Fully Tested) (AICodeKing on YouTube, September 24, 2026) (youtube.com), accessed
  21. Command Code's desktop app, tested: summary of AICodeKing's video (OpenClaw Database, September 24, 2026) (openclawdatabase.com), accessed
  22. Command Code in my team of AI tools (Md Zia Uddin, nzian.xyz, using it since March 2026) (nzian.xyz), accessed
  23. Claude plans and pricing (Anthropic, checked October 17, 2026) (claude.com), accessed
  24. Codex pricing (OpenAI, checked October 17, 2026) (learn.chatgpt.com), accessed
  25. OpenCode homepage (Anomaly, checked October 17, 2026) (opencode.ai), accessed
  26. OpenCode Go plans (Anomaly, checked October 17, 2026) (opencode.ai), accessed
  27. anomalyco/opencode repository, MIT licence (GitHub, checked October 17, 2026) (github.com), accessed
  28. Cursor pricing (Cursor, checked October 17, 2026) (cursor.com), accessed

Changelog

Changelog

  1. Page created from public reporting and public tests.

Recommended from GenGrades

  • Tool review

    T3 Code

    T3 Code is a free, MIT-licensed app from T3 Tools Inc., credited to Theo Browne's team, that runs the coding agents you already pay for, such as Claude Code, Codex, Cursor and OpenCode, side by side from a desktop, web or phone interface. First released on March 7, 2026, its GitHub repository had 28.1k stars on October 17, 2026.

  • Tool review

    ZCode

    ZCode is Zhipu AI's (Z.ai) free desktop coding-agent app for macOS, Windows and Linux, built around its GLM-5.3 model, with model use billed through the GLM Coding Plan from 18 USD a month. First released as Z Code in December 2025, it switched to Zhipu's own agent in June 2026; Z.ai said in August it had passed one million users.

  • Tool review

    VideoEnhancer

    VideoEnhancer is a web tool that upscales short MP4 clips to 1080p, 2K or 4K with the FlashVSR model and also compresses, face-swaps and generates video. After a 5-second free trial, 1080p costs 3 credits a second, $0.03 to $0.09 depending on how credits are bought. VideoEnhancer is our own product: it is made by the team behind GenGrades.

  • Tool review

    LayerGrab

    LayerGrab splits a flat image into up to 16 named, transparent layers over a filled-in background, on the web, in Figma or in Chrome, or pulls out only what you name. Each layer costs $0.10 on its $12 plan, $0.08 on Premium. No independent review exists yet. LayerGrab is our own product: it is made by the team behind GenGrades.

  • Tool review

    VideoTextRemover

    VideoTextRemover is a web tool that erases burned-in subtitles, captions and other on-screen text from MP4 and MOV videos and repaints what was behind them, up to 1080p. It bills per second processed, about $0.025 to $0.049, with 30 free seconds a month in some countries. VideoTextRemover is our own product: it is made by the team behind GenGrades.

  • Tool review

    MagicRemover

    MagicRemover is a browser tool that erases people, objects, text, logos and watermarks from photos, and subtitles from videos. Photos get 2 free removals a day, 2 more after sign-in, at about 1 megapixel; beyond that, credits start at $9.90 for 330. MagicRemover is our own product: it is made by the team behind GenGrades.