diff --git a/.gitignore b/.gitignore
index 198a7f66..79a9e247 100755
--- a/.gitignore
+++ b/.gitignore
@@ -136,4 +136,5 @@ zeroidc/target/
 __pycache__
 *.pyc
 *_source.tar.bz2
-snap/.snapcraft
\ No newline at end of file
+snap/.snapcraft
+tcp-proxy/tcp-proxy
diff --git a/tcp-proxy/Makefile b/tcp-proxy/Makefile
new file mode 100644
index 00000000..af4e71e3
--- /dev/null
+++ b/tcp-proxy/Makefile
@@ -0,0 +1,7 @@
+CXX=$(shell which clang++ g++ c++ 2>/dev/null | head -n 1)
+
+all:
+	$(CXX) -O3 -fno-rtti -o tcp-proxy tcp-proxy.cpp
+
+clean:
+	rm -f *.o tcp-proxy *.dSYM
diff --git a/tcp-proxy/README.md b/tcp-proxy/README.md
new file mode 100644
index 00000000..5af078a2
--- /dev/null
+++ b/tcp-proxy/README.md
@@ -0,0 +1,4 @@
+TCP Proxy Server
+======
+
+This is the TCP proxy server we run for TCP tunneling from peers behind difficult NATs. Regular users won't have much use for this.
diff --git a/tcp-proxy/tcp-proxy.cpp b/tcp-proxy/tcp-proxy.cpp
new file mode 100644
index 00000000..d5735198
--- /dev/null
+++ b/tcp-proxy/tcp-proxy.cpp
@@ -0,0 +1,317 @@
+/*
+ * ZeroTier One - Network Virtualization Everywhere
+ * Copyright (C) 2011-2016  ZeroTier, Inc.  https://www.zerotier.com/
+ *
+ * This program is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+// HACK! Will eventually use epoll() or something in Phy<> instead of select().
+// Also be sure to change ulimit -n and fs.file-max in /etc/sysctl.conf on relays.
+#if defined(__linux__) || defined(__LINUX__) || defined(__LINUX) || defined(LINUX)
+#include <linux/posix_types.h>
+#include <bits/types.h>
+#undef __FD_SETSIZE
+#define __FD_SETSIZE 1048576
+#undef FD_SETSIZE
+#define FD_SETSIZE 1048576
+#endif
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include <stdint.h>
+#include <unistd.h>
+#include <signal.h>
+
+#include <map>
+#include <set>
+#include <string>
+#include <algorithm>
+#include <vector>
+
+#include "../osdep/Phy.hpp"
+
+#define ZT_TCP_PROXY_CONNECTION_TIMEOUT_SECONDS 300
+#define ZT_TCP_PROXY_TCP_PORT 443
+
+using namespace ZeroTier;
+
+/*
+ * ZeroTier TCP Proxy Server
+ *
+ * This implements a simple packet encapsulation that is designed to look like
+ * a TLS connection. It's not a TLS connection, but it sends TLS format record
+ * headers. It could be extended in the future to implement a fake TLS
+ * handshake.
+ *
+ * At the moment, each packet is just made to look like TLS application data:
+ *   <[1] TLS content type> - currently 0x17 for "application data"
+ *   <[1] TLS major version> - currently 0x03 for TLS 1.2
+ *   <[1] TLS minor version> - currently 0x03 for TLS 1.2
+ *   <[2] payload length> - 16-bit length of payload in bytes
+ *   <[...] payload> - Message payload
+ *
+ * TCP is inherently inefficient for encapsulating Ethernet, since TCP and TCP
+ * like protocols over TCP lead to double-ACKs. So this transport is only used
+ * to enable access when UDP or other datagram protocols are not available.
+ *
+ * Clients send a greeting, which is a four-byte message that contains:
+ *   <[1] ZeroTier major version>
+ *   <[1] minor version>
+ *   <[2] revision>
+ *
+ * If a client has sent a greeting, it uses the new version of this protocol
+ * in which every encapsulated ZT packet is prepended by an IP address where
+ * it should be forwarded (or where it came from for replies). This causes
+ * this proxy to act as a remote UDP socket similar to a socks proxy, which
+ * will allow us to move this function off the rootservers and onto dedicated
+ * proxy nodes.
+ *
+ * Older ZT clients that do not send this message get their packets relayed
+ * to/from 127.0.0.1:9993, which will allow them to talk to and relay via
+ * the ZT node on the same machine as the proxy. We'll only support this for
+ * as long as such nodes appear to be in the wild.
+ */
+
+struct TcpProxyService;
+struct TcpProxyService
+{
+	Phy<TcpProxyService *> *phy;
+	int udpPortCounter;
+	struct Client
+	{
+		char tcpReadBuf[131072];
+		char tcpWriteBuf[131072];
+		unsigned long tcpWritePtr;
+		unsigned long tcpReadPtr;
+		PhySocket *tcp;
+		PhySocket *udp;
+		time_t lastActivity;
+		bool newVersion;
+	};
+	std::map< PhySocket *,Client > clients;
+
+	PhySocket *getUnusedUdp(void *uptr)
+	{
+		for(int i=0;i<65535;++i) {
+			++udpPortCounter;
+			if (udpPortCounter > 0xfffe)
+				udpPortCounter = 1024;
+			struct sockaddr_in laddr;
+			memset(&laddr,0,sizeof(struct sockaddr_in));
+			laddr.sin_family = AF_INET;
+			laddr.sin_port = htons((uint16_t)udpPortCounter);
+			PhySocket *udp = phy->udpBind(reinterpret_cast<struct sockaddr *>(&laddr),uptr);
+			if (udp)
+				return udp;
+		}
+		return (PhySocket *)0;
+	}
+
+	void phyOnDatagram(PhySocket *sock,void **uptr,const struct sockaddr *localAddr,const struct sockaddr *from,void *data,unsigned long len)
+	{
+		if (!*uptr)
+			return;
+		if ((from->sa_family == AF_INET)&&(len >= 16)&&(len < 2048)) {
+			Client &c = *((Client *)*uptr);
+			c.lastActivity = time((time_t *)0);
+
+			unsigned long mlen = len;
+			if (c.newVersion)
+				mlen += 7; // new clients get IP info
+
+			if ((c.tcpWritePtr + 5 + mlen) <= sizeof(c.tcpWriteBuf)) {
+				if (!c.tcpWritePtr)
+					phy->setNotifyWritable(c.tcp,true);
+
+				c.tcpWriteBuf[c.tcpWritePtr++] = 0x17; // look like TLS data
+				c.tcpWriteBuf[c.tcpWritePtr++] = 0x03; // look like TLS 1.2
+				c.tcpWriteBuf[c.tcpWritePtr++] = 0x03; // look like TLS 1.2
+
+				c.tcpWriteBuf[c.tcpWritePtr++] = (char)((mlen >> 8) & 0xff);
+				c.tcpWriteBuf[c.tcpWritePtr++] = (char)(mlen & 0xff);
+
+				if (c.newVersion) {
+					c.tcpWriteBuf[c.tcpWritePtr++] = (char)4; // IPv4
+					*((uint32_t *)(c.tcpWriteBuf + c.tcpWritePtr)) = ((const struct sockaddr_in *)from)->sin_addr.s_addr;
+					c.tcpWritePtr += 4;
+					*((uint16_t *)(c.tcpWriteBuf + c.tcpWritePtr)) = ((const struct sockaddr_in *)from)->sin_port;
+					c.tcpWritePtr += 2;
+				}
+
+				for(unsigned long i=0;i<len;++i)
+					c.tcpWriteBuf[c.tcpWritePtr++] = ((const char *)data)[i];
+			}
+
+			printf("<< UDP %s:%d -> %.16llx\n",inet_ntoa(reinterpret_cast<const struct sockaddr_in *>(from)->sin_addr),(int)ntohs(reinterpret_cast<const struct sockaddr_in *>(from)->sin_port),(unsigned long long)&c);
+		}
+	}
+
+	void phyOnTcpConnect(PhySocket *sock,void **uptr,bool success)
+	{
+		// unused, we don't initiate outbound connections
+	}
+
+	void phyOnTcpAccept(PhySocket *sockL,PhySocket *sockN,void **uptrL,void **uptrN,const struct sockaddr *from)
+	{
+		Client &c = clients[sockN];
+		PhySocket *udp = getUnusedUdp((void *)&c);
+		if (!udp) {
+			phy->close(sockN);
+			clients.erase(sockN);
+			printf("** TCP rejected, no more UDP ports to assign\n");
+			return;
+		}
+		c.tcpWritePtr = 0;
+		c.tcpReadPtr = 0;
+		c.tcp = sockN;
+		c.udp = udp;
+		c.lastActivity = time((time_t *)0);
+		c.newVersion = false;
+		*uptrN = (void *)&c;
+		printf("<< TCP from %s -> %.16llx\n",inet_ntoa(reinterpret_cast<const struct sockaddr_in *>(from)->sin_addr),(unsigned long long)&c);
+	}
+
+	void phyOnTcpClose(PhySocket *sock,void **uptr)
+	{
+		if (!*uptr)
+			return;
+		Client &c = *((Client *)*uptr);
+		phy->close(c.udp);
+		clients.erase(sock);
+		printf("** TCP %.16llx closed\n",(unsigned long long)*uptr);
+	}
+
+	void phyOnTcpData(PhySocket *sock,void **uptr,void *data,unsigned long len)
+	{
+		Client &c = *((Client *)*uptr);
+		c.lastActivity = time((time_t *)0);
+
+		for(unsigned long i=0;i<len;++i) {
+			if (c.tcpReadPtr >= sizeof(c.tcpReadBuf)) {
+				phy->close(sock);
+				return;
+			}
+			c.tcpReadBuf[c.tcpReadPtr++] = ((const char *)data)[i];
+
+			if (c.tcpReadPtr >= 5) {
+				unsigned long mlen = ( ((((unsigned long)c.tcpReadBuf[3]) & 0xff) << 8) | (((unsigned long)c.tcpReadBuf[4]) & 0xff) );
+				if (c.tcpReadPtr >= (mlen + 5)) {
+					if (mlen == 4) {
+						// Right now just sending this means the client is 'new enough' for the IP header
+						c.newVersion = true;
+						printf("<< TCP %.16llx HELLO\n",(unsigned long long)*uptr);
+					} else if (mlen >= 7) {
+						char *payload = c.tcpReadBuf + 5;
+						unsigned long payloadLen = mlen;
+
+						struct sockaddr_in dest;
+						memset(&dest,0,sizeof(dest));
+						if (c.newVersion) {
+							if (*payload == (char)4) {
+								// New clients tell us where their packets go.
+								++payload;
+								dest.sin_family = AF_INET;
+								dest.sin_addr.s_addr = *((uint32_t *)payload);
+								payload += 4;
+								dest.sin_port = *((uint16_t *)payload); // will be in network byte order already
+								payload += 2;
+								payloadLen -= 7;
+							}
+						} else {
+							// For old clients we will just proxy everything to a local ZT instance. The
+							// fact that this will come from 127.0.0.1 will in turn prevent that instance
+							// from doing unite() with us. It'll just forward. There will not be many of
+							// these.
+							dest.sin_family = AF_INET;
+							dest.sin_addr.s_addr = htonl(0x7f000001); // 127.0.0.1
+							dest.sin_port = htons(9993);
+						}
+
+						// Note: we do not relay to privileged ports... just an abuse prevention rule.
+						if ((ntohs(dest.sin_port) > 1024)&&(payloadLen >= 16)) {
+							phy->udpSend(c.udp,(const struct sockaddr *)&dest,payload,payloadLen);
+							printf(">> TCP %.16llx to %s:%d\n",(unsigned long long)*uptr,inet_ntoa(dest.sin_addr),(int)ntohs(dest.sin_port));
+						}
+					}
+
+					memmove(c.tcpReadBuf,c.tcpReadBuf + (mlen + 5),c.tcpReadPtr -= (mlen + 5));
+				}
+			}
+		}
+	}
+
+	void phyOnTcpWritable(PhySocket *sock,void **uptr)
+	{
+		Client &c = *((Client *)*uptr);
+		if (c.tcpWritePtr) {
+			long n = phy->streamSend(sock,c.tcpWriteBuf,c.tcpWritePtr);
+			if (n > 0) {
+				memmove(c.tcpWriteBuf,c.tcpWriteBuf + n,c.tcpWritePtr -= (unsigned long)n);
+				if (!c.tcpWritePtr)
+					phy->setNotifyWritable(sock,false);
+			}
+		} else phy->setNotifyWritable(sock,false);
+	}
+
+	void doHousekeeping()
+	{
+		std::vector<PhySocket *> toClose;
+		time_t now = time((time_t *)0);
+		for(std::map< PhySocket *,Client >::iterator c(clients.begin());c!=clients.end();++c) {
+			if ((now - c->second.lastActivity) >= ZT_TCP_PROXY_CONNECTION_TIMEOUT_SECONDS) {
+				toClose.push_back(c->first);
+				toClose.push_back(c->second.udp);
+			}
+		}
+		for(std::vector<PhySocket *>::iterator s(toClose.begin());s!=toClose.end();++s)
+			phy->close(*s);
+	}
+};
+
+int main(int argc,char **argv)
+{
+	signal(SIGPIPE,SIG_IGN);
+	signal(SIGHUP,SIG_IGN);
+	srand(time((time_t *)0));
+
+	TcpProxyService svc;
+	Phy<TcpProxyService *> phy(&svc,false,true);
+	svc.phy = &phy;
+	svc.udpPortCounter = 1023;
+
+	{
+		struct sockaddr_in laddr;
+		memset(&laddr,0,sizeof(laddr));
+		laddr.sin_family = AF_INET;
+		laddr.sin_port = htons(ZT_TCP_PROXY_TCP_PORT);
+		if (!phy.tcpListen((const struct sockaddr *)&laddr)) {
+			fprintf(stderr,"%s: fatal error: unable to bind TCP port %d\n",argv[0],ZT_TCP_PROXY_TCP_PORT);
+			return 1;
+		}
+	}
+
+	time_t lastDidHousekeeping = time((time_t *)0);
+	for(;;) {
+		phy.poll(120000);
+		time_t now = time((time_t *)0);
+		if ((now - lastDidHousekeeping) > 120) {
+			lastDidHousekeeping = now;
+			svc.doHousekeeping();
+		}
+	}
+
+	return 0;
+}