lan-direct

A skill and a browser MCP for coding agents behind a full-tunnel VPN — Claude Code and Codex.

Your VPN has to stay on, because that's how the agent reaches its API. But the app you're testing calls services that reject the VPN's foreign exit IP: a domestic API, a payment or SMS gateway, a bank endpoint. Turning the VPN off isn't an option — it cuts the agent off.

lan-direct gives the app an ISP-direct path while the agent stays on the tunnel. No admin, no routing-table edits, no VPN reconfiguration. How it works.

Install

Requires Node ≥ 20.6, plus PowerShell on Windows or bash and curl on Linux/macOS. No admin, no symlinks, no renaming.

From npm

npm install -g lan-direct
lan-direct install-skill                                             # or --codex
claude mcp add playwright-split --scope user "--" npx -y lan-direct browser

Then restart the agent. That's it — the same three lines on every platform.

npx -y lan-direct browser is the better MCP command however you install: it survives the checkout moving, which an absolute path into a clone does not.

Why "--" is quoted. On Windows claude resolves to claude.ps1, and PowerShell's parameter binder strips a bare -- before the script ever sees it. The CLI then reads -y as one of its own flags and fails with error: unknown option '-y'. Quoting protects it, and quotes are a no-op in bash — so the same line works in PowerShell, bash and cmd.

From a clone

Prefer this if you want to edit the skill or track it — one git pull updates everything, where npm needs npm update -g lan-direct && lan-direct install-skill because install-skill copies files rather than linking them.

Either way, your own host list at ~/.lan-direct-hosts is never touched.

1. Clone into your agent's skills directory. Both agents discover skills by scanning a directory for SKILL.md — they just scan different ones.

# Windows - Claude Code
git clone https://github.com/farshadmomo/lan-direct.git "$HOME\.claude\skills\lan-direct"

# Windows - Codex
New-Item -ItemType Directory -Force -Path "$HOME\.agents\skills" | Out-Null
git clone https://github.com/farshadmomo/lan-direct.git "$HOME\.agents\skills\lan-direct"

# Linux / macOS - Claude Code
git clone https://github.com/farshadmomo/lan-direct.git ~/.claude/skills/lan-direct

# Linux / macOS - Codex
mkdir -p ~/.agents/skills
git clone https://github.com/farshadmomo/lan-direct.git ~/.agents/skills/lan-direct

2. Register the browser MCP:

powershell -NoProfile -ExecutionPolicy Bypass `
  -File "$HOME\.claude\skills\lan-direct\scripts\install.ps1"

The installer verifies the bypass actually works, registers the server against this checkout, and prints the equivalent ~/.codex/config.toml block for Codex. Safe to re-run — it replaces the existing entry rather than duplicating it.

On Linux/macOS there's no installer script; register it directly:

claude mcp add playwright-split --scope user "--" \
  node ~/.claude/skills/lan-direct/scripts/lan-browser.mjs --port 8900

Restart the agent. Then /lan-direct in Claude Code, or in either agent just describe the problem — the description frontmatter is the trigger, firing on phrasings like "test this while my VPN is on" or "the API rejects the VPN IP".

To update: git -C <that directory> pull. No reinstall needed.

git clone https://github.com/farshadmomo/lan-direct.git "$HOME\.claude\skills\lan-direct"
New-Item -ItemType Directory -Force -Path "$HOME\.agents\skills" | Out-Null
cmd /c mklink /J "$env:USERPROFILE\.agents\skills\lan-direct" "$env:USERPROFILE\.claude\skills\lan-direct"

git clone https://github.com/farshadmomo/lan-direct.git ~/.claude/skills/lan-direct
mkdir -p ~/.agents/skills
ln -s ~/.claude/skills/lan-direct ~/.agents/skills/lan-direct

git clone https://github.com/farshadmomo/lan-direct.git "D:\projects\lan-direct"
cmd /c mklink /J "$env:USERPROFILE\.claude\skills\lan-direct" "D:\projects\lan-direct"

Which hosts go via the ISP

Rules live in lan-hosts.txt and are suffix matches:

.ir             every .ir host
digikala.com    also seller.digikala.com, api.digikala.com, ...

Anything not matched takes the tunnel. List domestic hosts only — adding a foreign one sends it out through an ISP that may filter it.

The shipped list covers .ir wholesale plus the Iranian services that live on other TLDs: payment (zarinpal, payping, nextpay), SMS/OTP (kavenegar, melipayamak, farazsms, ippanel), hosting (parspack, hostiran, mizbanfa), maps (neshan), commerce (digikala, torob, basalam), education and jobs (quera, faradars, maktabkhooneh, 7learn, irantalent), AI (hoosha, gapgpt), and the analytics domains those sites embed (najva, yektanet). Every one was checked: all are unreachable over the tunnel and respond over the ISP.

Adding your own

Put them in ~/.lan-direct-hosts, one per line, same format:

# my own
aparat.com
internal.corp

That file is never touched by git pull, and is read each time the browser starts — no reinstall, no restart of anything but the browser. Rules from all three sources merge: shipped file, your file, and any --lan-hosts a,b passed at launch.

Getting a rule wrong fails in a way that looks like the site being down, so the proxy prints the route it chose for every connection. Check that before blaming the site — or ask directly, without starting anything:

$ lan-direct route hoosha.com youtube.com api.anthropic.com
24 rules in effect

  hoosha.com                       lan    ISP (bound)
  youtube.com                      vpn    tunnel
  api.anthropic.com                vpn    tunnel

What's in the box

Two halves, installed together:

The MCP is what makes browsing work at all — Chromium has no source-binding flag, so a browser can only reach the ISP through a proxy. Without it the skill covers backends and CLIs but not pages.

How it works

Per-socket source-address binding. A socket explicitly bound to the physical adapter's IP takes that adapter's default route instead of the tunnel's:

$ curl -s https://api.ipify.org                            # 203.0.113.7    (VPN exit)
$ curl -s --interface 192.168.1.20 https://api.ipify.org   # 198.51.100.42  (real ISP)

No admin rights, no routing-table edits, no VPN reconfiguration, nothing global. Because it's per-socket it is opt-in by construction — a process that doesn't ask for it cannot be affected. That's the property that keeps the agent on the tunnel.

This is deliberately not route add. A static route is global: it would pull every process off the tunnel for that destination, the agent included, and it needs admin.

A server process has no --interface flag, so there are two ways to give it one:

The Node trap

If your backend is Node, HTTP_PROXY will not work and will fail silently.

Native fetch runs on Node's built-in undici, which ignores proxy environment variables on Node ≤ 23 and accepts no https.Agent. Verified on Node 22.20: plain env vars, NODE_USE_ENV_PROXY=1, and --use-env-proxy all do nothing — the proxy log stays empty while traffic goes out over the VPN. This hits Next.js, the openai SDK, and anything else on native fetch.

lan-bind.mjs patches net/tls in a preload, which reaches the bundled undici. That covers native fetch, the openai SDK, axios, node-fetch, got and plain http/https at once — no dependency, no proxy process, no HTTP_PROXY. NODE_OPTIONS propagates to forked workers, so it covers Next.js and Vite dev servers.

It skips loopback destinations on purpose: a socket sourced from the LAN IP cannot reach 127.0.0.1, so without that guard any app with a local Redis/Postgres/Mongo would die on EADDRNOTAVAIL.

Manual use

All paths below use $skill for wherever you installed it:

$skill = "$HOME\.claude\skills\lan-direct"    # Claude Code
$skill = "$HOME\.agents\skills\lan-direct"    # Codex

1. Detect and verify. Exits 0 only if the bypass actually works.

powershell -NoProfile -ExecutionPolicy Bypass -File "$skill\scripts\lan-detect.ps1"

bash ~/.claude/skills/lan-direct/scripts/lan-detect.sh    # Linux / macOS

Emits lanIp, gateway, lanDns, subnet, tunnelActive, vpnAdapter, directEgress, lanEgress, fakeIpDns, bypassWorks. It probes twice and compares, rather than assuming — and refuses to proceed if the bypass doesn't hold, so you never get a false pass.

2a. Node backend. Copy-paste as-is; only lanIp needs filling in:

$env:LAN_BIND     = "<lanIp>"
$env:NODE_OPTIONS = "--import file:///$($skill -replace '\\','/')/scripts/lan-bind.mjs"
npm run dev

Same thing from Git Bash, with the path written out:

LAN_BIND=<lanIp> \
NODE_OPTIONS="--import file:///C:/Users/YOU/.claude/skills/lan-direct/scripts/lan-bind.mjs" \
npm run dev

--import needs an absolute file:// URL with forward slashes on Windows — a bare Windows path fails.

2b. Anything else. Start the proxy:

node "$skill\scripts\lan-proxy.mjs" --bind <lanIp> --self-test    # run with the tunnel UP
node "$skill\scripts\lan-proxy.mjs" --bind <lanIp> --port 8899

Then launch the backend with HTTP_PROXY=http://127.0.0.1:8899, HTTPS_PROXY=http://127.0.0.1:8899, NO_PROXY=localhost,127.0.0.1,<subnet>.

See SKILL.md for stacks where HTTP_PROXY still isn't enough — notably PHP under Apache, where the service doesn't inherit your shell and libcurl never sees the variables.

Browsing both worlds at once

Neither path is a superset of the other. Measured on an Iranian ISP behind a foreign tunnel:

That last row is why the agent must never be moved off the tunnel — do it and the agent stops working entirely.

The browser MCP handles the mix in one session. Chrome is told to send everything to the proxy and makes no routing decision itself; the proxy reads the hostname out of each CONNECT (visible in plaintext, before TLS) and decides:

useBind(host) ? net.connect({ host, port, localAddress })   // ISP
              : net.connect({ host, port })                  // tunnel

There is no "use the VPN" code path. Binding is the only action taken — skip it and the OS routing table applies, which the tunnel owns. LAN is opt-in; the tunnel is what happens otherwise.

It is per connection, not per session. Two fetch calls from the same page can take different physical routes — measured, on one page in one tab:

Every proxy log line names the route it chose:

CONNECT  lan  hoosha.com:443
CONNECT  vpn  youtube.com:443

The browser's proxy runs on 8900, deliberately not the 8899 used for backend testing above. The MCP reuses whatever is already listening on its port, so sharing the number would let a bind-everything backend proxy silently capture the browser and push all of its traffic through the ISP — breaking exactly the sites the tunnel exists for.

Contents

Requires Node ≥ 20.6 and curl (built into Windows 10+, standard elsewhere). No dependencies.

Development

npm test          # node:test, no install step - there are no dependencies

34 tests. The routing predicate gets the most attention because it's the only logic here whose failure is silent — a wrong match sends traffic out the wrong physical network with no error and no log line. Both mutations it exists to catch are verified to fail the suite: a naive endsWith (so evilhoosha.com matches hoosha.com) and a missing loopback guard.

The preload is tested in a real child process with LAN_BIND set to 192.0.2.1 — an address the machine can't own, so a bound socket fails instantly. That's what proves the patch is active rather than quietly doing nothing.

What CI can't cover: lan-proxy.mjs --self-test, which needs a live tunnel and a physical adapter. Run it by hand with the VPN up. Changes are listed in CHANGELOG.md.

One gotcha if you're on an old runtime: Node 20.6.0 can't run the suite — every assertion passes and then the test runner never exits. That's a bug in that release's runner, not in lan-direct, which runs fine on it. Any current 20.x or 22 is fine.

Notes

• The proxy listens on 127.0.0.1 only. Binding wider would publish an open proxy to the LAN.
• lan-direct route <host> shows which network a host would leave by, using your merged rule list. Faster than starting the browser to find out.
• The .ps1 files are deliberately pure ASCII: Windows PowerShell 5.1 reads BOM-less UTF-8 as ANSI, and a stray em dash breaks the parser. CI enforces this.
• Under fake-IP DNS (sing-box, clash), pass --dns <lanDns>. dns.setServers() doesn't affect dns.lookup(), which goes through the OS resolver the tunnel owns.
• pkill -f lan-proxy from Git Bash does not kill a Windows node.exe. Use the PowerShell teardown in SKILL.md.

Verified against

nekoray / sing-box TUN on Windows 11, Node 22.20, Intel Wi-Fi — tested end to end on a real Next.js 16 app calling Neshan (Iranian maps) and AvalAI (LLM gateway), both of which reject foreign IPs outright:

The browser MCP was tested the same way: a domestic AI platform and YouTube opened back to back in one session, the domestic site's own CDN reporting the ISP address and country IR while YouTube rendered its German-locale shell over the tunnel.

The agent's own connection stayed on the tunnel throughout.

Adapter discovery is generic — Get-NetAdapter -Physical on Windows, a sysfs device symlink on Linux, networksetup on macOS — so WireGuard, OpenVPN, WARP and similar should work. But lan-detect verifies empirically rather than assuming, and exits non-zero when it can't.

Platform status. Windows is the tested path, end to end. lan-detect.sh has been run on Linux (Ubuntu 24.04) and returns the correct interface, gateway, DNS, routes and exit codes — but on a machine with no tunnel of its own, so the bypass-verified branch is unproven there. macOS is written from the documented behaviour of networksetup/ipconfig/netstat and has not been run at all. Both refuse rather than guess, so the failure mode is a clear error, not silent VPN traffic. Reports welcome.

One caveat specific to Linux: a bound source address does not by itself override the routing table the way it does on Windows, so a tunnel using policy routing may keep the traffic even when bound. That is exactly what the two probes detect — if it happens, lan-detect.sh says so and names the fix instead of reporting success.

Also unproven: lan-detect's "kill-switch detected" branch, which needs the VPN in a state the test machine couldn't produce.

License

MIT — see LICENSE.