Image: Appllama (GitHub)
Why it mattersAn indie builder with a coding agent can now reach a native-feeling screen on the first try. The screen still has to run in the simulator.
An indie developer tells a coding agent to build a settings screen. The result works, but it does not look like a real iPhone app. The spacing is off, the fonts are wrong, the back gesture is missing. A new repository on GitHub tries to close that gap.
Appllama released its appllama-skills repository on 13 August 2026. It has reached 4,517 GitHub stars, measured on the GitHub API on 4 November 2026. The repository is ranked 16 on trendshift.io's daily list. The licence is MIT.
What the repository ships
The repository has two agent skills, written for Claude Code, Codex, Cursor and other agents that read a skills folder.
The first skill, appllama-usage, is called the research engine. It tells the agent how to pull design references, flows and paywall screens through Appllama's own MCP server at mcp.appllama.io/mcp. That server needs an Appllama Pro account.
The second skill, appllama-app-design-skill, is the build skill. It tells the agent to turn a short brief into Expo and React Native code. The code has to match Apple's Human Interface Guidelines. That means semantic colours, native controls, push versus replace navigation, sheets, overlays, and springs that feel right on a phone.
The repository names its own rule for quality. Every screen the agent writes must pass a simulator verification loop. The agent runs the screen in an iOS simulator. The simulator takes the screenshot. A screenshot the agent drew from imagination does not count.
What sits behind it
The MCP the research skill calls is an index of real iOS apps. Appllama's own site states 84,800 screens from 1,840 top-grossing iOS apps, with new apps added each week. The company lists Amazon, Discord, Duolingo and the maker of ChatGPT among its customers. The MCP itself is paid. The two agent skills in the repository are free and open.
The limits worth naming
The repository is 83 days old, with 10 commits on its main branch. The author account on GitHub is called Appllama. The trademark notice in the readme names Antmind Ventures. The MIT licence covers the skill files only, not the MCP data.
The research skill without the MCP is a playbook. The build skill without the research skill is an opinionated screen generator with no references. A reader without the Pro account can still run the build skill from a brief. The agent then outputs a screen that follows Apple's Human Interface Guidelines. That is a lower bar than matching a named reference app. It is also where most agent output falls short today.
Think about a founder or a freelance developer writing a mobile app with an AI agent. The agent now has to run its own screen in a simulator before calling it done. That removes the most common failure mode. A paid design database then turns a native-feeling screen from a lucky draw into a routine output. Neither half is new on its own. Putting them in one install is the point.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
