porchlight

A small CLI + terminal app and AI skill that shares the command-line tools and MCP servers on your computer with MCP clients like Claude, over HTTPS.

porchlight turns df -h, a backup script or a local MCP server into tools that a remote client can call. Nothing is reachable until a person at your computer approves the client.

Features

• One command per tool: porchlight tools add disk -- df -h. Everything after -- is the command, word for word, with no shell.

• You approve every client: a client that asks to connect gets nothing until you type porchlight's code on its device.

• Links per audience: put your home tools on one address and your work tools on another. Each MCP server gets its own.

• Safe by default: commands that hand input to a shell are refused, every tool can be switched off, and every call is logged.

• Agent-friendly: every command takes --json, has examples in --help, and the skill teaches agents the workflow.

• No account needed: OpenTunnel gives you a public address, or bring your own domain (tunnels).

Related MCP server: cloud-to-local

Installation

$ curl -fsSL https://github.com/mahali00/porchlight/releases/latest/download/install.sh | sh
$ npx skills add mahali00/porchlight

macOS and Linux. The installer checks the download's checksum, and its build attestation when the GitHub CLI (gh) is installed. The second line is optional: it gives your coding agent the skill.

How it works

1. Share a tool. Wrap any program. A {name} becomes an input the client fills in:

$ porchlight tools add disk --read-only -d "Free disk space" -- df -h
$ porchlight tools add weather --read-only -d "Weather for a city" -- curl -s "wttr.in/{city}?format=3"
$ porchlight servers add files -- npx -y @modelcontextprotocol/server-filesystem ~/Documents
$ porchlight servers import          # lists the MCP servers Claude, Cursor and VS Code already use

A command that deletes or overwrites things, like rm {path}, is never marked read-only, and clients are told it's destructive. When a client chooses its input, a person has to confirm it in a terminal. No input can contain .., so rm ~/Downloads/{name} can't reach outside ~/Downloads.

2. Start sharing. porchlight opens the tunnel and keeps running in the background:

$ porchlight start
$ porchlight links list
● tools        https://porchlight.k2x9.opentunnel.xyz/tools/mcp
  disk, weather · 2 of 2 tools on · no clients

3. Approve the client. Add the address to your MCP client as a remote MCP server. It opens a page that asks for a code. porchlight shows that code on your computer: in a notification, in the app under Clients, on the home page, or with porchlight approve. Type it on the page, whether the client runs on this computer, a phone or another laptop:

$ porchlight approve
Code for Claude on tools: EEXR-YFWK
Type it on the device that's connecting. It expires in 9:41.

A wrong code keeps the page open and says how many tries are left; after five, the request ends. Only a person at your computer can see the code. Agents and scripts can do everything else.

TL;DR:

• Tools are programs you already trust, run as you, without a shell. Each input is one argument.

• A link is an address plus the tools on it. Change a link's tools and its clients approve again.

• porchlight logs shows every call, tools disable turns any tool off, and clients revoke takes access away.

The CLI

Every action in the terminal app (porchlight with no command) is also a command:

Full reference: CLI. Run porchlight <command> --help for examples.

The skill

npx skills add mahali00/porchlight installs a skill that teaches coding agents to set porchlight up and share tools safely. It's plain Markdown, so it doubles as the manual:

• Workflow: the steps an agent follows to share something.

• CLI: every command, option and exit code.

• Configuration: porchlight.json, including typed inputs, groups, environment and tool rules.

• Tunnels: OpenTunnel, or your own domain with cloudflared or a reverse proxy.

• What porchlight does not guarantee: what approval protects, and what it doesn't.

Development

$ cargo fmt --check
$ cargo clippy --all-targets --locked -- -D warnings
$ cargo test --locked

See CONTRIBUTING.md and AGENTS.md.

License

MIT