dfinity · raymondk · Oct 2, 2026
@@ -5,7 +5,7 @@ sidebar:
   hidden: true
 ---
 
-Developer tools are used to create, manage, and interact with canisters. ICP provides tooling across several categories: command-line tools, Motoko, canister development kits (CDKs), client libraries, testing tools, browser-based IDEs, and Candid tooling.
+Developer tools are used to create, manage, and interact with canisters. ICP provides tooling across several categories: command-line tools, Motoko, canister development kits (CDKs), client libraries, testing tools, browser-based IDEs, AI app builders, and Candid tooling.
 
 ## Command-line tools
 
@@ -164,6 +164,14 @@ Limitations:
 - Projects are limited to 5 MB and 2 canisters
 - ICP Ninja is not a replacement for icp-cli for production workflows
 
+## AI app builder
+
+### Caffeine
+
+[Caffeine](https://caffeine.ai) builds and hosts apps on ICP from a chat description. Its cloud agent writes a Motoko backend and a React frontend, builds them, and hosts the app on the Internet Computer; you publish it to a live URL from the Caffeine web app. No local toolchain is required.
+
+The optional [Caffeine CLI](https://www.npmjs.com/package/@caffeineai/cli) clones a project's source so you can edit it on your machine and upload it back as a new draft, and the hosted Caffeine MCP server lets an AI coding agent work with your Caffeine projects. See [Drive Caffeine from your agent](../guides/ai-coding-agents.md#drive-caffeine-from-your-agent) and the [Caffeine help center](https://help.caffeine.ai/hc/en-us).
+
 ## Candid tools
 
 ### didc

@@ -25,6 +25,21 @@ ICP has a set of [ICP skills](https://skills.internetcomputer.org): structured k
 
 **Learn more:** [AI coding agents](../guides/ai-coding-agents.md)
 
+## Ship an app without a toolchain
+
+**You want to:** Get a working app running on ICP without installing anything.
+
+[Caffeine](https://caffeine.ai) builds apps on ICP from a chat description. You describe what you want, and Caffeine's cloud agent writes the app (a Motoko backend and a React frontend), builds it, and hosts it on the Internet Computer. You publish it to a live URL from the Caffeine web app. Nothing runs on your machine.
+
+Caffeine suits a first prototype, or an app you intend to keep changing through chat. When you want to control the build yourself, choose the language, or use your own frontend stack, follow the [Quickstart](quickstart.md) and build with icp-cli instead.
+
+**Start with:** [caffeine.ai](https://caffeine.ai): describe your app and watch it build.
+
+**Then explore:**
+
+- [Drive Caffeine from your agent](../guides/ai-coding-agents.md#drive-caffeine-from-your-agent): work with your Caffeine projects from Claude Code, Cursor, ChatGPT, or another MCP client
+- [Caffeine help center](https://help.caffeine.ai/hc/en-us): Caffeine's own guides and reference
+
 ## Backend development
 
 **You want to:** Write canister logic: store data, call APIs, run scheduled tasks.

@@ -102,10 +102,28 @@ ICP skills are available without authentication:
 | Skill zip bundle | `https://skills.internetcomputer.org/.well-known/skills/{name}/SKILL.zip` |
 | Skills discovery index | `https://skills.internetcomputer.org/llms.txt` |
 
+## Drive Caffeine from your agent
+
+[Caffeine](https://caffeine.ai) builds and hosts apps on ICP from a chat description; see [Ship an app without a toolchain](../getting-started/choose-your-path.md#ship-an-app-without-a-toolchain). Its hosted MCP (Model Context Protocol) server lets the agent you already use work with your Caffeine account: list your projects, start a Caffeine build session, watch it progress, and continue a conversation that is running on Caffeine, without switching tabs.
+
+The connector drives Caffeine's cloud agent. It does not read or write code on your machine, and it does not send your project's source to your AI tool. Agents that can run shell commands (Claude Code, Cursor, Codex) can ask it for the local workflow instead: the Caffeine CLI clones the project, you edit and validate it locally, and the CLI uploads the result to Caffeine as a new draft. For the code itself, the agent should load the [`caffeine-app`](https://skills.internetcomputer.org/.well-known/skills/caffeine-app/SKILL.md) skill alongside `writing-motoko`.
+
+The server is at `https://mcp.caffeine.ai/mcp`, over Streamable HTTP with OAuth. In Claude Code:
+
+```bash
+claude mcp add --transport http --scope user caffeine https://mcp.caffeine.ai/mcp
+```
+
+Any client that supports remote MCP servers takes the same URL. The first time the agent calls a Caffeine tool, it opens your browser to sign in at caffeine.ai. To confirm the connection, ask the agent to list your Caffeine projects.
+
+:::note[Alpha]
+The Caffeine MCP server is in alpha, and its tool surface can change between releases, so this page does not list the tools. Setup for ChatGPT, Claude Desktop, Cursor, VS Code, Codex, and Perplexity, plus troubleshooting, is in Caffeine's [Setting up the Caffeine MCP server](https://help.caffeine.ai/hc/en-us/articles/49663075113236-Setting-Up-the-Caffeine-MCP-Server).
+:::
+
 ## Next steps
 
 - [skills.internetcomputer.org](https://skills.internetcomputer.org): browse all available ICP skills
 - [Developer tools](../developer-tools/index.md): icp-cli, CDKs, and other tools in the ICP toolchain
 - [Quickstart](../getting-started/quickstart.md): deploy your first canister with icp-cli
 
-<!-- Upstream: informed by dfinity/icskills — README.md, skills/*/SKILL.md -->
+<!-- Upstream: informed by dfinity/icskills — README.md, skills/*/SKILL.md; help.caffeine.ai — Setting Up the Caffeine MCP Server (article 49663075113236, updated 2026-06-04) -->