Fly.io

More setup than Railway or Render, but gives fine-grained control over region, scaling, and resources.

Install flyctl and log in

$ brew install flyctl   # macOS
$ fly auth login

Initialize the app

From the repo root:

$ fly launch --no-deploy

Accept the generated fly.toml. Pick a region close to you.

Set your secrets

$ fly secrets set \
>   MCP_TRANSPORT=streamable-http \
>   MCP_PORT=8080 \
>   PARADEX_ENVIRONMENT=prod

Configure and deploy

Confirm the [[services]] block in fly.toml listens on port 8080 with health check path /health:

1 [[services]]
2   internal_port = 8080
3   protocol = "tcp"
4 
5   [[services.tcp_checks]]
6     grace_period = "10s"
7     interval = "30s"
8     restart_limit = 0
9     timeout = "5s"
10 
11   [[services.http_checks]]
12     interval = "10s"
13     grace_period = "5s"
14     method = "GET"
15     path = "/health"
16     protocol = "http"
17     timeout = "2s"

Then deploy:

$ fly deploy

Your endpoint will be https://<app-name>.fly.dev/mcp.

Connect your AI client

Paste the URL into your AI client’s MCP settings. See the Remote MCP quickstart for client-specific instructions.

Fly.io keeps one VM running by default (no cold starts). Use fly scale count 0 to pause billing, fly scale count 1 to resume. fly logs streams live logs.

$	fly secrets set \
>	MCP_TRANSPORT=streamable-http \
>	MCP_PORT=8080 \
>	PARADEX_ENVIRONMENT=prod

1	[[services]]
2	internal_port = 8080
3	protocol = "tcp"
4
5	[[services.tcp_checks]]
6	grace_period = "10s"
7	interval = "30s"
8	restart_limit = 0
9	timeout = "5s"
10
11	[[services.http_checks]]
12	interval = "10s"
13	grace_period = "5s"
14	method = "GET"
15	path = "/health"
16	protocol = "http"
17	timeout = "2s"