Claude Code

These notes tell you how to install Claude Code under Harvard's official auspices, and have it direct bill to our lab funding.

1. Install Claude Code.

Go to claude.com/download and follow the instructions.

You won't be able to run it yet. You need an API key. That's next.

2. Apply to HUIT for your "app" API key.

Go to: portal.apis.huit.harvard.edu/register-an-app.

Your username in the upper right is a drop-down menu; select Apps.

Press the + NEW APP button in the upper right.

Provide the following information to the form:

App name: <LastName>_FAS
Example for me: Eddy_FAS
Description: <YourName> - FAS faculty fund - HUIT customer B40802 - monthly limit $400
Example for me: Sean Eddy - FAS faculty fund - HUIT customer B40802 - monthly limit $400
Owner: select me <your_email_address>
Example for me: me <seaneddy@fas.harvard.edu>
APIs: select AI Services - AWS Bedrock API by pressing the Request button.

When done, click the SAVE button in the lower right.

Approval is automatic, if HUIT's script sees our valid HUIT customer number.

This sets up monthly billing to the lab's startup fund. Our startup fund is a nigh-nonrenewable resource right now. Use it wisely. The HUIT script is supposed to see the $400 monthly limit and honor it, but I'm not yet sure that it does. I have been tracking my Claude session usage to make sure I'm not burning through money too outrageously.

Once you're approved, you will have access to a HUIT web page titled <LastName>_FAS for your Claude billing account. This page shows you your API key, unique to you. This is what you need to provide to Claude to enable access.

TREAT THE API KEY LIKE A PASSWORD. NEVER EXPOSE IT TO ANYONE. NEVER EXPOSE IT IN A GIT-MANAGED FILE.

3. Configure your environment.

There are a couple of ways to do this. I prefer to do it with a Claude-specific configuration file in my home directory: Claude reads .claude/settings.json on startup. Put the following in this file to start:

{
  "env": {
    "ANTHROPIC_BEDROCK_BASE_URL": "https://go.apis.huit.harvard.edu/ais-bedrock-llm/v2",
    "CLAUDE_CODE_USE_BEDROCK": "1",
    "CLAUDE_CODE_SKIP_BEDROCK_AUTH": "1",
    "ANTHROPIC_API_KEY": "<your_key_here>",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
  },
}

Obviously replace <your_key_here> with your API key.

Remember, DO NOT expose this configuration file to anyone.

The other way to do this is to set environment variables:

export ANTHROPIC_BEDROCK_BASE_URL=https://apis.huit.harvard.edu/ais-bedrock-llm/v2
export ANTHROPIC_API_KEY=<your_key_here>
export CLAUDE_CODE_SKIP_BEDROCK_AUTH=1
export CLAUDE_CODE_USE_BEDROCK=1
export CLAUDE_CODE_ATTRIBUTION_HEADER=0

I don't do it this way because I maintain my .bashrc under git control, and I use the same .bashrc on a wide array of different machines, not all of which are under my control. You can put these shell commands in a secret file somewhere that you execute as a shell script before you start Claude; I did it that way for a while but found it more tedious than using .claude/settings.json.

4. First time with Claude Code

cd to the directory you want to work in.

Run it at the command line: claude

Put it in a sandbox, so it only accesses files under the current directory (plus its own startup files in .claude/, I suppose) and doesn't go rummaging around elsewhere: /sandbox. Select auto-allow mode.

Select your model: /model. Select from the menu. I use Opus 4.8. Fable will generally refuse to work on comp bio code.

Off you go - ask it to work on something!

More information

These notes are an abbreviated version of notes from a "Generative AI for Scholarship" nanocourse that Chris Stubbs gave. His course is online and gives more detail than I have:

Setting up Anthropic API access

The Claude Code CLI

Anthropic also provides notes on getting started.