Open-source local AI coding tools · Your files stay on your machine · Explore on GitHub ↗
SFI Agentby Shah Fahad Inc.
Read docsInstall SFI Agent ↗
Quick start

Install once. Connect
your own tools.

Follow these steps on the computer where your code and development tools live.

1. Before you begin

Use a computer you control, with a supported Node.js runtime and a network connection. Keep your projects backed up and confirm you trust the source of any installer before running it.

Security note: piping a remote script directly to a shell executes whatever that URL serves. If you prefer to inspect it first, download install.sh, review it, then run it locally.

2. Install SFI Agent

On Linux or macOS, run the installer from a terminal:

curl -fsSL https://agent.shahfahadinc.com/install.sh -o sfi-agent-install.sh
less sfi-agent-install.sh
bash sfi-agent-install.sh

For a direct one-line install, if you have reviewed and trust the script:

curl -fsSL https://agent.shahfahadinc.com/install.sh | bash

On Windows, open PowerShell and use the published PowerShell installer after reviewing it:

irm https://agent.shahfahadinc.com/install.ps1 | iex

Installer behavior depends on the current release and operating system. Check the output and use the supported command listed by the installer.

3. Start the local control center

sfi-host --help
sfi-host --web

Open http://localhost:5180 on the same computer. Select only the project folders you intend to expose to your AI client.

4. Authorize a workspace

  1. Add or select a project directory in the local control center.
  2. Review allowed tools and permission mode.
  3. Keep secrets, private keys, credentials, and unrelated folders outside the allowed scope.
  4. Verify that the server can read an ordinary source file before attempting write operations.

5. Connect an AI client

Choose a connection method supported by your client. Local clients may connect to a local server; remote web clients generally require a reachable HTTPS endpoint and may require a compatible MCP transport and authentication configuration.

See AI client integrations and tunnel setup. The exact MCP URL is the endpoint shown by your running host configuration—not necessarily the public website homepage.

Do not expose the control-center dashboard publicly just to make MCP work. Publish only the MCP endpoint, and use authentication or an access policy appropriate to the tools being exposed.