Fline Skills MCP Server

fline skills

Skills that teach your AI coding app (Claude Code, Codex, Cursor and the other apps that read Agent Skills) how to put your project online with fline. For Claude Code, the repository is also a plugin that connects fline's MCP server.

Install

The Claude Code plugin brings fline's MCP server with it: type /mcp, choose fline, then Authenticate, and approve fline in your browser. Installed any other way, the skills carry no server; connect fline as the fline docs describe, or ask your AI app to, which loads fline-connect.

You need a fline account. fline is in private beta: see Getting your account.

What it sends, and where

The skills are instructions; on their own they send nothing anywhere. What leaves the computer goes through fline's MCP server at https://mcp.fline.sh/mcp, and only when the person asks for something fline does:

• The project folder. scripts/pack.sh packs a copy of it as a .tar.gz, leaving out .git, node_modules, what the project's .gitignore ignores, and .env and key files, and the agent uploads it to fline's storage. fline keeps that copy so it can build and run the app. The original stays in the folder.
• Parts of the code, to an AI model made by OpenAI, so fline can work out how to build and run the app. Passwords and keys saved on fline never go to it.
• Tool calls. Project names, settings names, web addresses and the other arguments of fline's tools. Passwords and keys are typed by the person on a private fline page, not sent through the chat. The exception is a key the agent already holds, one it generated or one the person pasted into the chat: set_env sends that value to fline, which stores it and never shows it again.
• What comes back. Project state, build output and the app's recent log lines, which hold whatever the app prints.

Sign-in is OAuth in the browser; the plugin holds no key. Disconnecting in the fline dashboard takes effect within 30 seconds. Details: privacy policy, terms, what fline does today.

Skills

They fall into three groups: using fline (connect, launch, database, domain, diagnose), preparing a project for it (prepare, build-plan), and making it safe to share (share-check).

Why skills when the MCP server has prompts

fline's server ships seven prompts with these procedures. In Claude Code an MCP prompt is a slash command the person types; the model has no tool to fetch one, so the server's instructions point at procedures the agent cannot open. A skill is loaded by the model itself when its description matches the task.

Skills name fline's tools without a prefix (create_project, not mcp__fline__create_project): the prefix differs between a plugin-provided server and one the person added.

Layout

.claude-plugin/plugin.json        the plugin, including fline's MCP server
.claude-plugin/marketplace.json   a marketplace holding that one plugin
skills/<name>/SKILL.md            one directory per skill (Agent Skills format)
skills/fline-launch/scripts/      pack.sh, run by the agent, never read
tests/                            scenarios and the projects they run on

Testing

• claude plugin validate --strict . checks the manifests.
• bash skills/fline-launch/scripts/pack.sh <folder> runs on the bash 3.2 and bsdtar that macOS ships, and on GNU tar.
• tests/scenarios.md holds six scenarios with what a passing answer does, run by a fresh agent with and without the skills. tests/make-fixtures.sh writes the two seeded projects they run on: a Node shop and a Flask tracker.

Where this comes from

Every fact here restates fline's own documentation, in particular for-ai.md. This repository is published from fline's main repository; a change made only here is replaced by the next publish.

License

MIT