DirectLink is a signaling-assisted peer-to-peer platform. It coordinates login, authorization, server registration, discovery, and connection setup; application data flows directly between endpoints over encrypted channels.
The platform has a hosted coordination service plus endpoint software that runs on customer machines. The same DirectLink foundation supports two product pairs: general private services through MTCode Server and Portal, and managed GPU clusters through MTGPU Server and its clients: the Remote-GPU extension and remotegpu-cli.
DirectLink coordinates identity and connection setup; endpoint traffic is encrypted and direct whenever possible.
Note: Click to view the MTGPU diagram, which is built on DirectLink.
Hosted by MTCodeAI.com for authentication, authorization, registration, service discovery, and connection signaling. It does not relay application data.
MTCode Server registers ordinary TCP services. MTGPU Server registers managed GPU hosts and dedicated SSH endpoints.
MTCode Portal maps authorized services to local ports. The Remote-GPU extension or remotegpu-cli maps authorized GPU servers and opens SSH terminals.
When a user signs in, the client retrieves only the services or GPU hosts that user is authorized to access. When the user connects, DirectLink signaling helps both endpoints establish a direct peer-to-peer path through NAT where possible. The endpoint protocol remains encrypted end to end.
DirectLink connects successfully for the great majority of users. To push the success rate as high as possible, DirectLink applies a multi-layered NAT-traversal strategy that adapts to different network configurations — it can establish direct connections even through many symmetric-NAT routers. Even so, some restrictive environments — typically a strict corporate firewall combined with certain types of symmetric NAT — can still block a direct connection. In those cases a connection may not be established; an administrator can set up a lightweight relay server, as described below.
For the rare networks where a direct path cannot be established, you can provide your own fallback. mtcode-relay, included in the MTCode Server and MTGPU Server packages, is a lightweight relay server that is easy to host. Run it on a VPS with a public IP or inside your private network behind a NAT router; it needs only one inbound port opened on both the router and firewall. When an administrator runs mtcode-relay under their account, it becomes an optional relay for every server and client on that account. mtcode-relay rejects any unauthorized connection.
Whenever a server and a client cannot reach each other directly, DirectLink automatically routes that connection through your relay — with no configuration or action from users. Direct connections are always preferred and used whenever possible; the relay steps in only when a direct path fails, and the whole handoff is completely transparent to users. For how to run it, including the one-port setup and the UDP and TCP modes, see the mtcode-relay manual.
Every connection between a server and its clients is protected by TLS encryption. When MTGPU Server, MTCode Server, or mtserver-cli registers with the platform, it generates its own certificate and private key locally on the server’s computer. The private key is created there and stays there — it is never uploaded, transmitted, or stored anywhere else.
Because no one else ever holds the key — not even the MTCode platform — no third party can decrypt the traffic between your clients and your servers. Encryption is truly end to end: your data can be read only at the two endpoints you control.
This protection holds even when a connection is routed through a relay. A relay forwards only the already-encrypted stream — it holds no private key and cannot read a single byte of your data.
Reverse-proxy and tunneling tools such as ngrok and Cloudflare Tunnel route traffic through an intermediate relay server. That extra hop can add latency, limit throughput, create bandwidth costs, and put a third-party provider in the traffic path. They also work by publishing a public endpoint: the service becomes reachable from the entire internet by anyone with the address unless access controls are added at the relay. And because the relay typically terminates TLS for HTTPS endpoints, the provider handles the traffic in plaintext unless the service supplies its own end-to-end encryption.
Mesh VPNs such as Tailscale connect entire devices into a private overlay network — but a reachable network is not the same as visible services. Nothing lists what is available: each user must already know which machine hosts a service and connect with the right address and port. And when you want to share only specific services — not the entire computer — across multiple administrators, careful access-control configuration and coordination are often required. Those access rules belong to the network's administrator, not to the person whose computer joins: under the default policy a joined device is reachable on every port, and beyond the device's own firewall the owner's controls are limited to blocking all incoming connections or leaving.
With MTCode DirectLink, users need one app — MTCode Portal — and one account. Authorized services from every administrator appear automatically in one interface, ready to use without network configuration. Access is granted to individual services, not entire computers — and neither server nor client computers join any network or expose anything. Connections are established only for the specific services in use.
Start with the product that matches your need: a managed GPU development host, or a private service published through DirectLink.