Command Code

Command Code is a commercial AI coding agent sold by subscription, distributed as a terminal program on the npm registry under the package name command-code and, since September 2026, as a beta desktop application. Its terms of service and privacy policy, both last updated 20 September 2026, identify the vendor as "Langbase, Inc. d/b/a Command Code," with Delaware governing law: Command Code is the trading name of the San Francisco AI infrastructure company Langbase, whose founder and CEO is Ahmad Awais.^[31][32] The product was announced publicly on 25 February 2026 together with a self-announced $5M seed round.^[2][3] It positions itself as a harness tuned for open-weight models rather than as a model of its own, and its distinguishing feature is a component the company calls taste-1, which it says learns a developer's coding preferences from accepted, rejected and edited suggestions and applies them to later work.^[2][4]

Almost every quantitative claim the company publishes about its scale, efficiency and rank is self-reported, and several of the numbers on its home page changed substantially during September 2026, so each one is attributed to the company where it appears below. The two measurements it does not control are its npm download counts and the position of the commandcode.ai entry on OpenRouter's public app board, both given below with the date they were read. The taste-1 description is also disputed: an independent teardown of the published package, reproduced for this article against the current release, finds no model by that name in the shipped code.

Company and launch

The npm registry entry for command-code was created on 7 August 2025, when version 0.0.1 was published; the package then spent roughly six months in alpha and beta prereleases.^[1] The company's own changelog dates the first versioned release, v0.1.0, to 26 November 2025 and labels it "Initial release - AI-powered coding assistant in your terminal."^[5]

The public launch came on 25 February 2026, with a blog post headed "Introducing Command Code: the coding agent that learns your taste" that opened: "We've raised $5M to launch the first coding agent that can continuously learn your coding taste."^[3] A longer launch essay at commandcode.ai/launch, published under Awais's byline, gives the round's composition: PWV, the fund of GitHub co-founder Tom Preston-Werner, first invested at pre-seed and led the seed round, alongside funds the post lists as Firststreak, ENEA, Mento, Banyan, Alumni and AltaIR, plus "many a16z scouts" and a group of angel investors.^[2] The named angels on that page and on the company's about page include Preston-Werner, Apple CFO Luca Maestri, Cloudflare CTO Dane Knecht, Supabase CEO Paul Copplestone, Replit CEO Amjad Masad, Snyk founder Guy Podjarny, Socket CEO Feross Aboukhadijeh, Resend CEO Zeno Rocha, Arcjet CEO David Mytton, Logan Kilpatrick of Google, and Theo Browne of T3 Chat.^[2][6] Every detail of the round, including the amount, comes from the company's own pages. No filing naming Langbase appears in the SEC's EDGAR full-text search index, read 1 October 2026, and searches of the technology trade press turned up no reported coverage of the round, only machine-generated funding-database entries and aggregator pages whose text could not be retrieved.^[34]

The same launch essay sets out the company's lineage, and the legal paperwork settles it. Awais writes that frustration with AI developer tooling led him to leave a VP job and found Langbase, describes Langbase's products as Pipes and Memory, and presents Command Code as the applied form of "over a year of research" done there.^[2] The corporate trail matches. Langbase's own site still operates as a serverless AI developer platform, carries a banner reading "Command Code is live now," gives Command Code a section of its home page, and signs its footer "© 2026 Langbase, Inc. 2261 Market St #5698, San Francisco, CA 94114."^[30] Command Code's footer gives the identical address under the name Command Code.^[4] The GitHub organisation that hosts Command Code's public repositories, CommandCodeAI, was created on 20 November 2023 and still contains Langbase's open source projects including BaseAI, langbase-sdk and LangUI, and its older status repository is still described as "⌘ Langbase," alongside a newer commandcode-status repository described as "⌘ Command Code."^[7] The npm package's keyword list includes "langbase."^[1] The terms of service and the privacy policy both name the contracting party as "Langbase, Inc. d/b/a Command Code," with Delaware governing law, so Command Code is a trading name of Langbase rather than a separate company.^[31][32] Awais's personal site states the same thing in its page title, which reads "Founder & CEO of CommandCode.ai f/k/a Langbase."^[33]

Awais is listed as founder and CEO; the about page lists eleven people in total, and four of the npm package's publishing maintainers (asharirfan, saqibameen, ahmadawais, ahmadbilaldev) correspond to names on that page.^[6][1]

What it ships

The CLI requires Node.js 22 or newer and installs four executable names: cmd, cmdc, commandcode and command-code. The package readme notes that on Windows the short alias is cmdc, because cmd is already the built-in Windows command shell.^[1] The npm license field is UNLICENSED and the published tarball ships only built output; the package's development dependencies include javascript-obfuscator.^[1]

The desktop app puts chat, a file tree, a Git working-tree view, a terminal, a local browser preview and a plan view in one window, and bundles the agent runtime so the CLI is not required. Its documentation carries a beta notice telling users to review the agent's actions, back up their work and watch model costs, and the announcement post describes a roadmap of "computer use, sandboxes, cloud computers, and a couple of secret projects."^[11][10] A "Design mode" in the browser pane lets a user select an element in a local page and hand it, with page context, to the chat; the docs warn that it should be used only with local development pages the user trusts.^[11]

The Provider API exposes Command Code's routed model pool through OpenAI Chat Completions, OpenAI Responses and Anthropic Messages endpoints at api.commandcode.ai/provider/v1/, plus a systemone endpoint for the TypeSafe AI decision model Jev.^[9] A zero-data-retention mode is enabled per session with the CMD_ZDR=1 environment variable or an x-cmd-zdr: 1 header, and the documentation says a ZDR request fails with HTTP 422 rather than falling back to a provider that retains data.^[13]

Models and how access is sold

Command Code routes work to third-party models rather than serving one of its own. The CLI's generated model reference listed 86 model identifiers across 21 vendor sections when read on 1 October 2026, including Alibaba, Anthropic, DeepSeek, Google, inclusionAI, Meituan, Meta, MiniMax, Moonshot AI, NVIDIA, OpenAI, Poolside, Sakana AI, StepFun, Tencent, Thinking Machines, TypeSafe, xAI, Xiaomi, Z AI and a "Stealth" section for unreleased models.^[14] The default model on that date was deepseek/deepseek-v4-flash, an open-weight model, which is consistent with the product's open-models positioning.^[14]

Access is sold three ways at once. Subscriptions bundle a monthly allowance of model credits; a separate Provider plan is metered pay-as-you-go; and a beta BYOK path lets a user attach their own provider key, with requests going from the user's machine straight to that provider. The BYOK menu is documented as covering "150+ providers out of the box" with endpoint and model list prefilled, and also accepts an existing ChatGPT Plus/Pro or GitHub Copilot subscription, or a local server such as Ollama, vLLM or LM Studio.^[12]

Prices read from the pricing page on 1 October 2026, all per month and all before a card processing fee the page notes separately:

Source: commandcode.ai/pricing, read 1 October 2026.^[15] The request-volume column is the company's own approximation and the page says it "depends on chosen models." Top-up credits are sold at the underlying API rate with no markup, roll over and do not expire, according to the same page.^[15] The GOAT plan, the tier the site pushes hardest, was introduced on 5 August 2026.^[16] All plans are advertised with "no training on your code" and local-only taste storage.^[15]

Some models are offered free for a period, including models the pricing page lists only by codename (Space Bunny Alpha and Pixel Canary, both described as stealth previews), and the 1 October 2026 changelog entry retiring Pixel Canary dates the end of its preview to 30 September 2026.^[15][5]

taste-1

taste-1 is the component the company presents as its research bet, and it is the one part of the product its documentation describes in mechanical detail. The company says taste-1 is not a large language model: it generates no text, and instead takes a signal (an accept, a reject, an edit, or a correction diff) and returns a decision about whether that signal is a real preference, which category it belongs to, and a confidence between 0 and 1.^[17] The architecture is described as "meta neuro-symbolic": a neural half that forms an intuition from extracted signals, a symbolic rule layer that fixes the category set, scores confidence and flags contradictions rather than overwriting an earlier preference, and a "meta" loop that re-scores old learnings as new evidence arrives.^[17]

The documented pipeline has four steps, and the company is explicit that two of them are ordinary LLM calls on a model the user picks: an LLM turns prompts and diffs into structured signals (step 1), taste-1 classifies them (step 2), an LLM writes the result out as readable Markdown (step 3), and the CLI stores it (step 4).^[17] The stored output is plain Markdown in the project at .commandcode/taste/taste.md, with per-line confidence scores, and it can be read, edited, linted with npx taste lint or reset.^[17][2] Profiles are shareable as "taste packages" via npx taste push and npx taste pull, scoped to a project, to the user's machine, or to a remote profile on the company's Studio.^[17]

The launch essay describes the learning objective as a "meta neuro-symbolic RL objective" in which a user's edits and feedback become reward signals, states that the approach was chosen over fine-tuning, retrieval and prompt injection, and calls it "an early-stage research direction."^[2] The same page adds a caution rare in the company's material: "We're being careful not to overstate this. The system learns patterns, not intentions."^[2] The documentation says taste-1 is on introductory pricing, included on every plan at no extra charge within unspecified "generous limits," and that the user pays only for the LLM work at steps 1 and 3.^[17] The company has published no paper, model card or weights for taste-1, and no third-party evaluation of it has been published, so every figure attached to it is the company's own. The description itself has been challenged by an independent reading of the shipped code, covered in the next section.

The documentation itself draws a comparison to TypeSafe AI's Jev, which it calls "the clearest public example" of the same class of non-generative decision model, and Command Code ships Jev as a selectable model in headless mode and on the Provider API.^[17][14]

Scrutiny of the taste-1 description

On 19 September 2026 an independent teardown of the published package appeared at cmd.safzan.dev under the title "There is no model called taste-1," written by Safzan Pirani, who also published the longer version as a GitHub gist. It pins its analysis to command-code@1.58.0, gives the file it read (dist/cli.mjs, 2,635,431 bytes), publishes the SHA-256 of that file, de-minifies the bundle with prettier and lists the exact grep commands a reader can run to check each claim.^[35][36] Its central finding is that the string taste-1 appears five times in the shipped code, in FAQ copy, a /taste help string, a billing string and a status-bar literal, and nowhere as a model identifier, routing entry or catalogue record, while the bundle carries roughly eighty real model identifiers that do have those things.^[35]

The teardown's account of what the feature actually does is a five-step pipeline: git show mines before-and-after line pairs from up to 200 commits; compileTasteContext() concatenates the diff hunks into a Markdown prompt ending in a block headed "Instructions for Taste Learning" with eight style guidelines; the result is posted to api.commandcode.ai/alpha/generate on whatever model the user already selected; the model writes Markdown to .commandcode/taste/taste.md; and renderTasteSection2() pastes that file into the system prompt on later turns. "Nothing in it computes embeddings or updates weights," the page says, "and no step performs reinforcement learning."^[35] It also notes that the status-bar badge is produced by a line that appends " · taste-1" only when the user is not on their own API key, so the label disappears while the feature behaves identically.^[35]

Those findings were reproduced for this article against the current release rather than the one the teardown read. Read on 1 October 2026, the unpkg copy of command-code@1.58.0's dist/cli.mjs is 2,635,431 bytes with the SHA-256 the page publishes, and the current release command-code@1.73.2 behaves the same way: taste-1 occurs five times in the same four kinds of location, there is no model identifier for it, and runTasteLearningAgent, compileTasteContext, renderTasteSection2, the "Instructions for Taste Learning" prompt block and the alpha/generate endpoint are all present.^[37] The in-product strings the teardown quotes are also present verbatim in both bundles, including "Privacy: Taste processing runs on your codebase and stores learning data locally only" and "Command Code does not train on your code or store your code snippets."^[37] The teardown's objection to the first of those is that the storage claim holds, because the Markdown file is written locally, while the processing claim does not, because the compiled prompt containing verbatim source lines is posted to the vendor's API.^[35]

The teardown makes four further claims that this article did not attempt to verify beyond confirming that the named functions and constants exist in the bundle: that the taste extractor is the one file-reading path, among several, that does not consult the package's sensitive-filename denylist, so a secret committed once and later deleted can still be read out of history and uploaded; that taste learning also reads saved transcripts from Claude Code, Codex and Cursor off disk; that a device fingerprint hashing the user's git email, machine ID and MAC addresses is posted with a salt published inside the package itself; and that the server's "your CLI is out of date" error checks only whether a version header is present and well formed.^[35]

The company answered part of the objection the following day, and rewrote its documentation around the same time. On 20 September 2026 Awais published a long post, which the site marks as a 14-minute read, headed "taste-1 vs Jev: two decision models that are not LLMs," which asks "Is it an LLM? No. Then why does taste make LLM calls? Because a model that does not produce text needs one that does, on both sides of it," diagrams the four-step pipeline with LLMs at steps 1 and 3, and concludes that "LLM calls in the traffic are the design, not a contradiction of it."^[41] The post neither names the teardown nor addresses its two central findings, that the shipped bundle contains no model identifier for taste-1 and that the compiled prompt containing verbatim source lines is posted to the vendor's API; it does not use the words teardown, bundle or upload at all.^[41] It also introduces a new self-reported figure, calling taste-1 "one of the largest decision-model deployments in production" at "100T-token scale this year."^[41]

The documentation moved in the same direction on the same dates. The version of the Taste page read on 1 October 2026 states plainly that "taste-1 is not an LLM," sets out the four-step pipeline, tells the reader that "inspect the traffic and you will see LLM calls on the model you picked," and repeats the "design, not a contradiction" line.^[17] None of that material appears in the Internet Archive's capture of the same page from 7 September 2026, which described taste-1 only as "the core of our taste architecture" with no mention of LLM calls; all of it appears in the capture from 21 September 2026.^[38][39] The product itself was not changed: the FAQ string shipped in the current release still reads "taste-1 is our meta neuro-symbolic AI model with continuous reinforcement learning (RL)."^[37]

Community engagement with either the product or the criticism has been slight. The three Hacker News submissions about Command Code drew 3 points and no comments (6 May 2026), 6 points and two comments (a "Show HN" posted by the founder on 5 August 2026), and 3 points and no comments for the teardown (19 September 2026).^[40]

Harness engineering and the TEF benchmark

Command Code publishes a documentation series it calls Harness Engineering, whose premise is that "most of what gets reported as a model failing is really a model plus a harness failing."^[18] Two entries carry the numbers the company quotes most often.

"Tool Call Repairs," dated 3 May 2026, describes a validation-and-repair layer that rewrites malformed tool calls in flight. The company says four recurring mistakes (sending null for an optional field, emitting an array as a JSON string, wrapping a single argument in an object, and passing a bare string where an array was expected) account for most open-model tool-call failures across DeepSeek, GLM and Qwen models, and that repairs now run on every model it serves at "roughly 1M tool calls per 1T tokens, free on every plan." The post's headline claim, that the layer made DeepSeek V4 Pro beat Claude Opus 4.7 "6/10 times," is explicitly described as a result on the company's internal evals.^[19]

"The Shell Tool," dated 30 August 2026, is the source of the "token efficiency frontier" or TEF figures that appear on the home page. It is the company's own benchmark, not a third-party one. It grades 23 shell-tool capabilities across ten harnesses, assigns each capability an estimated token saving per million tokens of shell traffic, and reports that Command Code captures about 300K of roughly 306K savable tokens, about 98 percent, against 240K for Hermes Agent, 234K for OpenClaw, 219K for Grok Build, around 211K for Claude Code, 155K for Codex, 120K for Kilo Code, 83K for opencode, 63K for Cline and 40K for Pi.^[20] The method section is unusually candid about the limits: the nine open-source competitors were read by an AI model from pinned commits dated 28 to 30 August 2026, Claude Code's column was probed live because it ships no source and is called "the least certain column in the table," the capability prices are "modeled estimates from our production traces, not measurements," and the whole thing carries the disclaimer that "this benchmark and its analysis were produced by AI with little human review, and should be read that way."^[20]

The same post says the company planned to open-source the CLI "soon."^[20] As of 1 October 2026 it had not: the CommandCodeAI/command-code repository on GitHub contains only a readme and an issue tracker, with no license file and no source, and the npm package is published as UNLICENSED.^[7][1] The repository showed 4,072 stars, 359 forks and 358 open issues when read that day.^[7]

Release history

The company shipped continuously and in small increments: its changelog counted 411 releases on 1 October 2026, split as 387 CLI, 24 Desktop and 80 pre-release entries, and npm listed 466 available versions of command-code.^[5][1] Only the releases that changed what the product is are listed here.

Version numbers in the changelog do not line up exactly with npm publish times for the pre-launch releases: npm's first stable 0.9.0 was published on 8 January 2026, while the changelog dates v0.9.0 to 25 December 2025.^[1][5] Dates in the table above are the changelog's except where an npm version is named.

Adoption

Two measurements of Command Code's use exist that the company does not control.

The first is npm. The command-code package was downloaded 221,966 times in the 30 days ending 29 September 2026 and 58,278 times in the week ending the same date.^[21] Monthly totals climbed from 4,429 in March 2026 to 61,785 in May, 177,386 in July and 307,378 in August before easing to 212,339 in September.^[21] These are download events, not installs or users: the CLI updates itself in the background unless run with --no-auto-update, so each user generates many downloads over time.^[5]

The second is OpenRouter's public app board, which ranks applications by the tokens they route through OpenRouter and counts only apps that send attribution headers, so traffic a harness sends straight to a model provider is invisible to it. Read on 1 October 2026, Command Code sat 8th on the trailing day with 379B tokens, 8th on the trailing week with 2.2T, and 11th on the trailing month with 2.3T.^[22] That the weekly and monthly totals are nearly equal means almost all of its monthly OpenRouter volume arrived in that final week. Its OpenRouter app page that day gave 2.54T total tokens, a daily global rank of #8, "active since May 2026," and 35 models used; the top model in its last-30-days breakdown was the stealth model Space Bunny Alpha at 2.43T tokens, with GPT-5.6 Luna a distant second at 59.4B. Set against the 2.54T all-time total that is roughly 96 percent, although the two figures cover different windows.^[23] Space Bunny Alpha is one of the models Command Code lists as free on every plan while its preview lasts, so the bulk of its measured OpenRouter traffic came from requests its own customers paid nothing for.^[15]

Command Code also has no claimed app profile on OpenRouter. Claude Code, Cline, Kilo Code, Codex, Pi, Hermes Agent and OpenClaw each have a curated page with a written description; Command Code's listing is keyed to the bare URL commandcode.ai, as are those of Oh My Pi, DeepSeek Harness and Freebuff, and the descriptive text OpenRouter shows for Command Code, Oh My Pi and DeepSeek Harness is just that domain, while Freebuff supplies a tagline.^[22]

Figures the company reports about itself

The company's home page carried the following claims when read on 1 October 2026, none of which is independently verifiable: "100K+ developers," "40K+ paid customers," "30T scale," "130T+ tokens served," "99%+ cache hit rates, the best in the industry," tool call repairs at "about 1M repairs per 1T tokens," and the headline description "the best coding agent for open models."^[4] The same page claims "Code 10x faster. Reviews, 2x quicker. Bugs 5x fewer," and the launch essay supports those figures with a table of "correction loops" measured on the company's own task set, falling from between 2.9 and 4.2 edits without taste to between 0.2 and 0.5 edits after a month.^[4][2]

Several of these are recent. The Internet Archive's capture of the home page on 1 September 2026 shows a different headline ("Command Code with your taste," "the frontier coding agent") and a call to action reading "Join 29K+ developers." The 100K+ figure, the 40K+ paid-customer figure, the token totals and the open-models framing are all on the page read on 1 October 2026 but absent from that capture, so they were added during September 2026.^[24][4] Earlier milestone posts give the trajectory the company reported at the time: 100 billion tokens on 10 May 2026; "$1M run rate, 1 trillion tokens and 9K customers" on 30 May 2026, which the post dates to "24 days after public beta launch"; and 10,000 paying customers on 3 June 2026.^[25][26][27]

The "most used coding agent on OpenRouter" claim

On 20 July 2026 the company published a post headed "Command Code is the most used coding agent on OpenRouter," whose body reads: "Command Code is the most used coding agent on open router overtaking Claude Code, Kilo, OpenClaw, etc."^[28] The post's own category list, printed directly beneath that sentence, claims #1 in Programming App, #1 in Native App Builders and #1 in IDE Extensions, but #2 in Personal Agents, #2 in Coding Agents, #2 in Productivity and #2 in CLI Agents.^[28] A post four days earlier, on 16 July 2026, had claimed "#1 top trending agent," #3 in Coding Agents and #3 global rank.^[29]

The claim cannot be checked against the board as it stood on that date. The Internet Archive holds captures of openrouter.ai/apps from 1, 4, 12, 14, 17 and 28 July 2026, but the archived copies retrieved on 1 October 2026 contain only numbered placeholders, with no app names or token counts, so there is no archived record of the ordering in July 2026.^[42]

What can be checked is the present. On OpenRouter's reading of 1 October 2026, Command Code's own app page placed it #8 in Coding Agents and #8 in CLI Agents, behind Hermes Agent, Claude Code, Cline, Kilo Code, Pi, Oh My Pi (listed as omp) and Codex on the trailing week.^[23][22] The ranking is a volatile snapshot of one gateway's traffic, so neither the July claim nor the October reading describes overall market share. On the board the claim invoked, Command Code did not hold the top coding-agent position in any of the three windows on 1 October 2026.

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11npm registry. "command-code" package document (versions, publish times, maintainers, license, readme). Read October 1, 2026. registry.npmjs.org/command-code
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Ahmad Awais. "Command Code raises $5M to build the first coding agent that continuously learns your coding taste." Command Code launch essay. Read October 1, 2026. commandcode.ai/launch
3. ^1 ^2 ^3Team Command Code. "Introducing Command Code: the coding agent that learns your taste." February 25, 2026. commandcode.ai/...ing-agent-that-learns-your-taste
4. ^1 ^2 ^3 ^4 ^5Command Code. Product home page. Read October 1, 2026. commandcode.ai
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Command Code. "Changelog." Read October 1, 2026. commandcode.ai/changelog
6. ^1 ^2Command Code. "About Command Code." Read October 1, 2026. commandcode.ai/about
7. ^1 ^2 ^3GitHub API. Organization `CommandCodeAI` and repository `CommandCodeAI/command-code`. Read October 1, 2026. api.github.com/...command-code
8. ^1 ^2Team Command Code. "Introducing the Command Code Provider API." May 21, 2026. commandcode.ai/...ng-the-command-code-provider-api
9. ^1 ^2 ^3Command Code Docs. "Command Code Provider API." Read October 1, 2026. commandcode.ai/...provider
10. ^1 ^2 ^3Team Command Code. "Introducing the Command Code Desktop App." September 16, 2026. commandcode.ai/...ing-the-command-code-desktop-app
11. ^1 ^2 ^3Command Code Docs. "Command Code Desktop App." Read October 1, 2026. commandcode.ai/...desktop
12. ^1 ^2Command Code Docs. "BYOK." Read October 1, 2026. commandcode.ai/...byok
13. ^Command Code Docs. "Zero Data Retention." Read October 1, 2026. commandcode.ai/...zdr
14. ^1 ^2 ^3Command Code Docs. "Available models." Read October 1, 2026. commandcode.ai/...models
15. ^1 ^2 ^3 ^4 ^5Command Code. "Pricing." Read October 1, 2026. commandcode.ai/pricing
16. ^1 ^2Command Code. "Introducing the GOAT plan." August 5, 2026. commandcode.ai/...introducing-the-goat-plan
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Command Code Docs. "Taste." Read October 1, 2026. commandcode.ai/...taste
18. ^Command Code Docs. "Harness Engineering." Read October 1, 2026. commandcode.ai/...harness-engineering
19. ^Ahmad Awais. "Tool Call Repairs." Command Code Docs, Harness Engineering, May 3, 2026. commandcode.ai/...tool-call-repairs
20. ^1 ^2 ^3Ahmad Awais. "The Shell Tool." Command Code Docs, Harness Engineering, August 30, 2026. commandcode.ai/...shell-tool
21. ^1 ^2npm registry download counts API. Read October 1, 2026. api.npmjs.org/...command-code
22. ^1 ^2 ^3OpenRouter. "App & Agent Rankings." Read October 1, 2026. openrouter.ai/apps
23. ^1 ^2OpenRouter. Command Code app page. Read October 1, 2026. openrouter.ai/...https%3A%2F%2Fcommandcode.ai%2F
24. ^Internet Archive. Capture of commandcode.ai, September 1, 2026. web.archive.org/...commandcode.ai
25. ^Team Command Code. "Milestone: Command Code broke 100 billion tokens." May 10, 2026. commandcode.ai/...nd-code-broke-100-billion-tokens
26. ^Team Command Code. "$1M run rate, 1 trillion tokens and 9K customers in 24 days." May 30, 2026. commandcode.ai/...llion-tokens-and-9k-customers-in
27. ^Team Command Code. "Command Code crossed 10K paying customers in 30 days." June 3, 2026. commandcode.ai/...-10k-paying-customers-in-30-days
28. ^1 ^2Team Command Code. "Command Code is the most used coding agent on OpenRouter." July 20, 2026. commandcode.ai/...-used-coding-agent-on-openrouter
29. ^Team Command Code. "Command Code is the #1 trending agent on OpenRouter." July 16, 2026. commandcode.ai/...e-1-trending-agent-on-openrouter
30. ^Langbase. Company home page (footer: "© 2026 Langbase, Inc. 2261 Market St #5698, San Francisco, CA 94114"). Read October 1, 2026. langbase.com
31. ^1 ^2Command Code. "Terms of Service" (last updated September 20, 2026; "Langbase, Inc. d/b/a Command Code"). Read October 1, 2026. commandcode.ai/terms
32. ^1 ^2Command Code. "Privacy Policy" (last updated September 20, 2026). Read October 1, 2026. commandcode.ai/privacy
33. ^Ahmad Awais. Personal site, page title "Ahmad Awais - Founder & CEO of CommandCode.ai f/k/a Langbase". Read October 1, 2026. ahmadawais.com
34. ^U.S. Securities and Exchange Commission. EDGAR full-text search for "Langbase" (0 results). Read October 1, 2026. efts.sec.gov/...search-index
35. ^1 ^2 ^3 ^4 ^5 ^6Safzan Pirani. "There is no model called taste-1." September 19, 2026. cmd.safzan.dev
36. ^Safzan Pirani. "Command Code 1.58.0 taken apart: the taste subsystem end to end, every system prompt, tools, wire protocol, telemetry." GitHub gist. gist.github.com/...26170636512c0b50494d6a70acfece8d
37. ^1 ^2 ^3unpkg. `command-code@1.58.0/dist/cli.mjs` and `command-code@1.73.2/dist/cli.mjs` (published bundles). Read October 1, 2026. unpkg.com/...cli.mjs
38. ^Internet Archive. Capture of commandcode.ai/docs/taste, September 7, 2026. web.archive.org/...taste
39. ^Internet Archive. Capture of commandcode.ai/docs/taste, September 21, 2026. web.archive.org/...taste
40. ^Hacker News (via the Algolia API). Items 48031887, 49188656 and 49766313. Read October 1, 2026. hn.algolia.com/...49766313
41. ^1 ^2 ^3Ahmad Awais. "taste-1 vs Jev: two decision models that are not LLMs." September 20, 2026. commandcode.ai/...taste-1-vs-jev
42. ^Internet Archive. CDX index query for captures of openrouter.ai/apps, July to August 2026, and the retrieved captures of 12 and 17 July and 3 August 2026. Read October 1, 2026. web.archive.org/...cdx