Skip to main content

Run Claude Code on Radium

Point the Claude Code CLI at Radium's Anthropic-compatible Messages API. Setup script, manual configuration, model selection, and troubleshooting.

Run Claude Code on Radium

Claude Code sends Anthropic-compatible API traffic to whatever endpoint you point it at, and Radium serves an Anthropic-compatible Messages API, so moving Claude Code across is a change of base URL, model name, and key.

Your projects stay where they are. So do your prompts, your permissions, your local tools, and any MCP servers you have configured. Claude Code keeps running the session and Radium answers the model calls.

Supported models hal-1.0, clarke-1.0, and tycho-1.0. The examples on this page use hal-1.0, and any of the three can be substituted wherever it appears.

What changes

What you change

  • The API base URL
  • The model name
  • Your credentials
  • One settings file

What stays put

  • Your prompts and project context
  • Local tools and permissions
  • MCP servers and their configuration
  • Everything about how you work

Settings used throughout this guide

Setting Value
API base URL https://api.radium.cloud
Model hal-1.0, clarke-1.0, or tycho-1.0
Authentication Radium API key
Settings file ~/.claude/settings.json

Read this one twice Claude Code appends /v1/messages to the base URL on its own, so leave it off. Setting ANTHROPIC_BASE_URL to the full Messages endpoint is the most common way this goes wrong, and it returns a 404 every time.

Before you start

  • A Radium API key
  • Claude Code, installed and current
  • Node.js 18 or later, if you plan to run the setup script or the verifier
  • macOS, Linux, or a supported Windows environment such as WSL or Git Bash

Check what you have:

node --version
claude --version

If Claude Code is not installed:

npm install -g @anthropic-ai/claude-code

Setup with the script

Run this from the folder containing the integration package:

./01-setup-radium-api.sh \
  --api-base api.radium.cloud \
  --model hal-1.0

The script asks for your API key without printing it, and then it does five things:

  1. Normalises the base URL, so a trailing slash or a missing scheme does not break anything.
  2. Backs up your existing ~/.claude/settings.json with a timestamp.
  3. Preserves every setting already in the file that has nothing to do with Radium.
  4. Writes the base URL, the model, and the key.
  5. Restricts the file permissions to your user.

The backup lands beside the settings file, named like this:

~/.claude/settings.json.bak.20260807153000

Nothing you had configured is lost, and you can put it back by hand in one move.

Setup by hand

If you would rather not run a script against your machine, add these entries to ~/.claude/settings.json yourself:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.radium.cloud",
    "ANTHROPIC_MODEL": "hal-1.0",
    "ANTHROPIC_API_KEY": "YOUR_RADIUM_API_KEY"
  }
}

Keep any other settings that are already in the file, and keep the JSON valid. Then restrict access to it:

chmod 600 ~/.claude/settings.json

Shell environment variables work too, and they override the settings file, which is worth remembering when something behaves strangely later:

export ANTHROPIC_BASE_URL="https://api.radium.cloud"
export ANTHROPIC_MODEL="hal-1.0"
export ANTHROPIC_API_KEY="$RADIUM_API_KEY"

Verify the connection

Test Radium directly first, then test Claude Code. Doing it in that order tells you which of the two is at fault when something fails.

Step one, the endpoint on its own:

./02-test-radium-api.mjs --model hal-1.0

Step two, through Claude Code:

claude -p "Say only: radium model connected."

You should get back:

radium model connected.

Then open a project and work normally:

cd YOUR_PROJECT
claude

If step one passes and step two fails, the problem is in Claude Code's configuration and not in the endpoint.

Choosing a model

Primary model selection

Model Built for
hal-1.0 Higher-capability coding and agentic work
clarke-1.0 General-purpose and structured work
tycho-1.0 Faster, lighter-weight tasks

To change it, run the setup script again with a different model:

./01-setup-radium-api.sh \
  --api-base api.radium.cloud \
  --model clarke-1.0

Another backup is written before anything is overwritten.

A faster model underneath

Claude Code reaches for a smaller model to handle helper operations during a session, and those operations are the kind of work Tycho is built for. Running Hal as the primary with Tycho underneath puts each task on the model sized for it, and the session gets quicker in the places where capability was never the constraint.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.radium.cloud",
    "ANTHROPIC_MODEL": "hal-1.0",
    "ANTHROPIC_SMALL_FAST_MODEL": "tycho-1.0",
    "ANTHROPIC_API_KEY": "YOUR_RADIUM_API_KEY"
  }
}

Some Claude Code versions and some Radium deployments need this set explicitly, and others do not. Confirm the mapping with Radium before you rely on it in a managed deployment.

Bearer-token deployments

Radium deployments authenticate one of two ways. The default is an API key, and some deployments use a bearer token instead. Use whichever mode Radium supplied, and start with the API key when nobody has told you otherwise.

With the script:

./01-setup-radium-api.sh \
  --api-base api.radium.cloud \
  --model hal-1.0 \
  --auth bearer

By hand, swap the key variable for the token variable:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.radium.cloud",
    "ANTHROPIC_MODEL": "hal-1.0",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_RADIUM_API_KEY"
  }
}

Set one or the other. Setting both is a 401 waiting to happen.

When it does not work

404 on every request The base URL has a path on it. Use https://api.radium.cloud and let Claude Code add /v1/messages itself.

401 or 403 Check that the key belongs to the environment you are pointing at, check whether the deployment expects an API key or a bearer token, and check that only one of ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN is set. Restart Claude Code after any change.

Claude Code still appears to reach Anthropic Look at ~/.claude/settings.json and confirm the base URL, then look for shell environment variables overriding it, because they win. Close any open sessions and start a new one. If you are still unsure, run the direct verifier, which separates an endpoint problem from a configuration problem.

Model not found Use the exact identifier issued for your account. This integration supports hal-1.0, clarke-1.0, and tycho-1.0.

Tools or streaming behave differently Claude Code uses streaming, tool calls, and more than one model route in a single session. API compatibility covers the shape of the request and it does not guarantee that every model responds the same way, so run the workflows your team actually uses before a broad rollout.

Security and operations

  • Never commit ~/.claude/settings.json or any populated credential file.
  • The setup script stores your key locally in that settings file, so organisations that require short-lived credentials should use a managed key helper or a gateway instead.
  • Prompts, tool results, and attached project context are data sent to the configured endpoint, and they should be treated that way.
  • Use separate keys for development, staging, and production.
  • Test quality, tool behaviour, latency, and usage on repositories that look like your real ones.

Point it at a real project, run a day of work through it, and you will know.

Next

Back to top