What it costs

Running the template has two costs: the Railway server, and the Claude usage of your jobs.

Railway

The template needs Railway’s Hobby plan or above. Railway bills by use.

At idle, the three services use about 1.4 GB of memory in all:

Service	Memory at idle
Paperclip	about 0.9 GB
Postgres	about 0.37 GB
CloudCLI	about 0.17 GB

At Railway’s memory price when this manual was written, that is roughly $14 a month before runs. Check Railway’s current pricing.

Each job run adds a Claude Code process on Paperclip while it runs, and each chat you open adds one on CloudCLI.

Where it shows: each service’s Metrics tab in Railway, and your workspace’s usage page.

Claude

Each run is a Claude Code session on your own Claude subscription (Pro or Max). There is no separate bill: runs count toward your plan’s usage, like your own Claude Code sessions.

The Runner uses Sonnet by default. As a measure, the open-issues digest of the quickstart took:

Model	Tokens	Time
Sonnet	about 86,000 (10.5k input, 946 output, 74.7k cached)	12 seconds
Opus	about 150,500	not measured

Most of the tokens were cached. A longer skill, or one that reads more, uses more.

To change the model: open the Runner’s page in Paperclip (Agents, Runner) and pick another. Opus draws your plan’s limits down faster; a job that follows a written skill rarely needs it.

Where it shows:

• log in the chat: ask Claude “show the last log of the digest”. It shows each agent run’s tokens. See Runs, logs and cost.
• Paperclip: Agents, Runner, Runs, then a run. It shows input, output and cached tokens, the model and the duration.

Runs on a Claude subscription show no API cost. Look at the tokens.

Keep runs cheap

• One run per job. A skill written for a job ends by closing its task. A run that leaves its task open makes Paperclip start a second run just to close it.
• No backlog. Runs missed while the server was down are dropped, not replayed. See No overlap, no backlog.
• Pause what you don’t need. Ask Claude to pause a job; it stops running until you resume it.
• Codex jobs use your ChatGPT plan instead. See Codex.