How to Deploy an MCP Server or AI Agent from GitHub

What Unyly builds for each project type

You do not write a Dockerfile (though a repo Dockerfile is respected if present). Unyly detects the stack and generates one. For an MCP server on Node it installs deps and wraps the stdio entry (bin > main > index.js) with supergateway, exposing streamable HTTP at /mcp. Python is the same with a python:3.12 base. For a site it uses nginx:alpine for plain static, an npm build that ships dist/build/out, or a Node runtime for Next.js. Monorepos: set a Root Directory (web/, app/, …) or let autodetect find it.

# The generated Dockerfile for a Node MCP server (simplified)
FROM node:22-slim
WORKDIR /app
COPY package*.json ./
RUN npm install --omit=dev
COPY . .
RUN npm install -g supergateway
EXPOSE 3000
CMD supergateway --stdio 'node index.js' \
    --outputTransport streamableHttp --port 3000 --streamableHttpPath /mcp

Model keys, streaming and timeouts for AI apps

Unyly is not a model gateway. Your app calls the provider directly with your own key, set as a project env var (ANTHROPIC_API_KEY, OPENAI_API_KEY, …) — it is injected only into your container. The reverse proxy does not buffer responses, so SSE token streaming reaches the client as it is produced, and it holds a single request open for up to 10 minutes, which covers slow completions and multi-step agent tool loops. Containers get 384 MiB RAM and half a CPU on the current plan.

Deploying by talking to an AI agent

Unyly Deploy is exposed as an MCP server itself. Connect https://deploy.unyly.org/mcp in Claude, Cursor or ChatGPT (auth with a token from unyly.org/account/gateway, or Authorize via OAuth in claude.ai/ChatGPT) and the agent gets tools: deploy_project, list_projects, get_project, get_logs, redeploy, rollback, set_env, add_domain. So "deploy github.com/me/my-agent as an mcp and set ANTHROPIC_API_KEY" becomes a single instruction the agent carries out end to end.