CommandCode.ai

CommandCode.ai builds Command Code, a command-line AI coding agent centered on learning a developer’s “coding taste”: tooling choices, code structure, naming conventions and other recurring decisions. Ahmad Awais is its founder and CEO. He identifies the company as formerly Langbase, whose earlier products provided infrastructure for building and deploying AI agents with memory and tools.[2][1][3]

Command Code installs with npm i -g command-code and runs as cmd. Its defining feature is Taste, powered by taste-1. The company describes an architecture that turns accepted suggestions, rejected suggestions and subsequent edits into symbolic constraints that condition future LLM generation. The learning loop generates code using current preferences, observes the developer’s response, extracts or updates constraints, saves them and applies them to the next generation.[1][4]

In the launch announcement, Awais describes choosing a meta neuro-symbolic architecture after experimenting with fine-tuning, retrieval-augmented generation and rules in prompts. The company’s rationale was to separate a lightweight, continuously updated preference layer from the language model’s generation capability. It describes edits and feedback as reward signals for a reinforcement-learning objective, while calling the research direction early-stage. Awais gives its boundary plainly: “The system learns patterns, not intentions.” It cannot anticipate architectural decisions the developer has never shown it.[1]

Preferences remain visible and editable. Taste packages organize learned patterns into categories such as CLI conventions, TypeScript and architecture, each containing a taste.md file. Project packages live under .commandcode/taste/; personal packages live under ~/.commandcode/taste/. Developers can inspect, edit or reset learned preferences, and push, pull and compose profiles through Command Code Studio. Learning can be switched off at project or user scope; with it off, stated preferences remain in the current session without being saved for future use.[4][1]

Awais illustrates the intended behavior with a date CLI demonstration. He reports giving Claude Code and Command Code the same prompt after both had seen him build CLIs. Command Code’s output used TypeScript, Commander.js, a starting version of 0.0.1, a lowercase -v flag and ISO date formatting—choices he says it learned from previous sessions. He then transferred his CLI preferences between projects using npx taste push --all and npx taste pull ahmadawais/cli. This was the founder’s demonstration, rather than an independently reproduced comparison.[1]

The coding agent followed earlier work on AI infrastructure at Langbase. That history supplies context for the company’s move from agent-building primitives to continuous preference learning.[1][5]

Awais traces his own coding-agent work to July 2020, when, he recalls, Greg Brockman gave him early GPT-3 access and he built a CLI agent called clai. He later left a vice-president role to launch Langbase, attributing that decision to frustration with AI developer tooling.[1]

Langbase introduced PIPE, the Prompt Instructions Personalization Engine, in February 2024. A Pipe exposed a custom AI agent as an API, combining language models, tools and knowledge from datasets. Organization support followed in March, and a TypeScript SDK was released on July 31, 2024. These products established the agent infrastructure that Awais says preceded work on continuous learning.[5][1]

Another product, Command.new, turned prompts into multi-agent workflows with agent applications and infrastructure. The Langbase changelog records a Chai.new release on May 19, 2025 and a Chai.new-to-Command.new rename entry on July 23. In his personal account, Awais now calls Langbase and Command.new legacy products, says both are “getting acquired,” and says he and his team are fully focused on Command Code. His statement describes acquisitions in progress without identifying buyers, terms or completion dates.[5][3]

CommandCode’s development milestones span several releases. The changelog records a Taste Learning System on September 26, 2025; a platform launch with organizations, billing, authentication and Web Studio on September 27; Taste Learning V2 with dedicated learning-agent infrastructure on October 27; and a beta release on November 30. A December 15 entry records the launch page announcing the seed round.[5]

The Command Code team includes Saqib Ameen, Maedah Batool, Ahmad Bilal and Ashar Irfan as founding engineers, Omar Imtiaz in founding operations and Obaid Nadeem as design engineer. Langbase lists the same people in those roles. Command Code also lists Vipul Gupta as DX engineer and Naymur Rahman, Ali Arain and Shubham Yadav as software engineers.[2][6]

Command Code announced $5 million in seed funding in its launch material. Awais says PWV by Tom Preston-Werner, which had invested in the pre-seed, was leading the seed round. Named angel investors include Paul Copplestone, Amjad Masad, Guy Podjarny, Feross Aboukhadijeh and Zeno Rocha. The December 15, 2025 changelog milestone dates the announcement-page release, rather than establishing a financing closing date.[1][5]

The organization’s reported scale figures concern its Langbase platform. Langbase reports having helped more than 100,000 developers from more than 5,000 companies, processing 700-plus TB of vector memory and billions of agent runs each month. Command Code’s launch material separately reports 850 TB of AI memory and 1.2 billion monthly agent runs for Langbase. These undated company figures do not establish a growth sequence, a paying-customer count or Command Code adoption.[6][1]