Agent-first
When your project works, your AI looks at what it built and asks one question: what will you want to do with this every day? Then it builds those actions for real, tests them, and hands them to you as tools.
You describe what you want. Your AI does the work on your machine, shows you every command, and leaves you with something you can actually use tomorrow.
Sign in with your email, copy one line, paste it into your server. The agent installs itself and starts in the background. It survives restarts.
~30 secondsAdd your one Corypt MCP link to Claude. Say “build me a notes app” or “set up a VPN”. Your AI writes the code, runs it as a service, opens the port and checks it from the internet.
/srv/corypt/<project>The agent-first step designs the everyday actions for your project, builds them, checks they work and registers them. No MCP server written by hand.
tested on registerOpen any chat with the same link and use your tools. Share one tool with one person: they use it with their own link, and never see your server or your build access.
per tool, per personwget -qO corypt.sh https://build-mcp.corypt.com/corypt.sh &&
chmod +x corypt.sh &&
./corypt.sh install YOUR-CODEYour passwords, your photos, your wallets: the things you would rather keep on your own machine than in someone else’s cloud. Corypt builds them on your server once, and from then on you run them by just asking.
No dashboards to learn, no logins to remember. Open any chat, say what you need, and it is done on your server. Share one tool with one person, and they get that tool and nothing else.
Most AI tools stop at “the files are on your server”. Corypt does the parts that are usually skipped: running it for real, checking it from outside, writing down what it did, and handling your keys.
Build, deploy and use your tools through one MCP link. What it can do depends only on its token, so a scoped token makes a use-only link.
mcp.corypt.comImport a folder that already runs on your server. Corypt reads it and writes its INTRODUCTION.md, and never changes or moves a single file.
read-only importA free port picked for you, a service that survives reboots, and the firewall opened: ufw, firewalld, iptables or nftables. If your cloud has its own firewall, you get the exact rule to add.
service + portBefore “done”, Corypt opens the URL from Cloudflare’s network and, when needed, from volunteer servers in other countries. Works for you means works for everyone.
outside checkCorypt fetches what was built and reports what is really on it. It spots broken pages that still answer 200: tracebacks, app crash screens, gateway and nginx errors.
page checkA new project starts in draft: your AI writes the charter, the phases and the decisions before any code. Whoever opens the project next reads that plan and the written trail, and carries on.
draft, then devSave API keys once on the Integrations page, or log in with Cloudflare. Your AI asks Corypt for a key by name when a command needs it, and tells you if one has expired.
IntegrationsImprovement logs are off unless you switch them on in Settings. Switching them off deletes them right away.
off by defaultConnecting your own server costs nothing. And it stays yours. Corypt never receives your SSH keys or passwords, and never logs in to your machine. A small agent on your server reaches out to us; nothing ever reaches in. It only runs jobs that your own AI session asks for, signed with your own token.
Outbound only
The agent asks Corypt for work over normal HTTPS. Your server keeps zero open ports for us.
The whole agent is a single bash script using only bash, curl and jq. No hidden binaries. Read every line before you install it.
Each job is logged on your own server. Run corypt logs any time to see exactly what ran.
Delete a token in your dashboard and every link that uses it stops working right away. Scoped tokens limit what each link can do.
sudo systemctl stop corypt pauses it. sudo corypt uninstall removes it. Your projects stay in /srv/corypt.
Registered tools run through fixed handlers, not pasted shell text, which closes the classic injection gap.
The honest part: to build and run software, the agent works as root on your server. That is what makes it useful, and it is why you should only connect a server you are happy to let your AI manage. Use a fresh VPS for experiments, and keep your tokens private like passwords.
Already paying for a VPS? That is the only bill. Corypt does not charge you to connect it, build on it or use your tools.
Plenty of services can deploy code with AI. Very few leave you with something your chat can keep controlling, on hardware you own.
| Corypt | AI deploy platforms | Hand-written MCP server | |
|---|---|---|---|
| Runs on a server you own | Yes | Usually their cloud | Yes |
| Cost to connect your own server | Free | Not offered | Your time |
| Every project becomes chat tools | Automatic | No | You write each one |
| Share one tool with one person | Built in | No | You build auth |
| Checked from the internet before “done” | Built in | Varies | You check |
| A new chat continues where the last one stopped | Written trail on your server | Rarely | No |
| Open ports needed | None | None | Usually yes |
| Leave any time | Code stays on your disk | Export needed | Yes |
Still unsure? Connect a spare server for ten minutes and watch the log. That answers most of them.
Yes. Connecting your own Linux server costs nothing. You keep paying your server provider as usual; Corypt adds no bill on top.
No. We never get your SSH keys or passwords and you open no ports. The agent on your server connects out to Corypt and only runs jobs that come from your own AI session with your token. Every job is logged on your server.
Linux with bash and curl. If jq is missing, the installer adds it for you. It works on x86_64 and ARM64, and starts with systemd, OpenRC or cron, so it runs on a €2 VPS as well as on a big dedicated box.
Claude works out of the box. Any other assistant that supports MCP connectors can use the same links.
No. One Corypt MCP link does both: your AI builds and deploys with it, and uses your project tools and the tools other people shared with you. What a link can do depends only on its token. For a link that can use tools but never build, create a scoped token with just “Use project tools”. Older Build MCP and Projects MCP links keep working; they are the same connector now.
Yes. Ask it to import the folder. Corypt reads it, writes a short INTRODUCTION.md about it, and never changes or moves your files. Then your AI can deploy it, check it and give it chat tools like anything it built itself.
Nothing is lost. Corypt writes every command, and the reasoning behind it, into the project folder on your server. Open a new chat, say “continue this project”, and your AI reads that trail and carries on from where the last chat stopped.
On your server, in /srv/corypt/<project-name>. They are plain folders with plain files. If you ever stop using Corypt, they stay exactly where they are.
Run sudo systemctl stop corypt to pause it. Run sudo corypt uninstall to remove the service, its settings and the agent, then remove the server in your dashboard. Your projects in /srv/corypt are not touched.
Connect a server you already have, open Claude, and say what you want built. In a few minutes it is running on your machine, with tools waiting in your chat.