Run Claude Code with SAP Generative AI Hub Models

In Part 1 we built an Envoy Proxy that attaches an OAuth 2.0 JWT and rewrites request paths so they match SAP AI Core’s inference-endpoint schema for OpenAI models.

In this follow-up we will:

• Add support for Anthropic models to the Envoy configuration.
• Use the Anthropic Python SDK to call the Sonnet-4.5 model hosted in AI Core.
• Configure Claude Code so it can generate code with the same model.

Using the Anthropic Python SDK

Start the Envoy with our configuration from the Part 1 envoy-aicore.yaml:

$ envoy -c envoy-aicore.yaml

Next, run a slightly modified example that calls a Sonnet-4.5 model :

from anthropic import AnthropicBedrock

client = AnthropicBedrock()

response = client.messages.create(
    model="d7ed726e32394096",
    max_tokens=1024,
    tools=[
        {
            "type": "bash_20250124",
            "name": "bash"
        }
    ],
    messages=[
        {"role": "user", "content": "List all Python files in the current directory."}
    ]
)
print(response.content)

Two changes compared with the standard sample:

• We instantiate AnthropicBedrock, because AI Core provides Anthropic models through AWS Bedrock.
• We pass the AI Core Deployment ID (d7ed726e32394096) instead of a model name.

For convenience we will supply the base URL via an environment variable rather than hard-coding it:

$ ANTHROPIC_BEDROCK_BASE_URL=http://localhost:8100 \
  uv run -w anthropic["bedrock"] test-anthropic.py

If you run the script now you will receive:

anthropic.NotFoundError: Error code: 404

Envoy logs reveal why:

[2025-12-14T16:57:56.549Z] "POST /model/d7ed726e32394096/invoke HTTP/1.1" 404 - 198 0 130 130 "-" "AnthropicBedrock/Python 0.75.0" ...

AI Core expects the path /v2/inference/deployments, not /model.

Insert a new route into envoy-aicore.yaml:

- match:
    prefix: "/bedrock/model"
  route:
    prefix_rewrite: "/v2/inference/deployments"
    auto_host_rewrite: true
    cluster: aicore
    timeout: 300s
  request_headers_to_add:
  - header:
      key: "AI-Resource-Group"
      value: "default"
    append_action: ADD_IF_ABSENT

Here we:

• Use the /bedrock prefix in the base URL to keep the config organized.
• Replace the /bedrock/model prefix with the required AI Core path.
• Add the mandatory AI-Resource-Group header.
• Forward the request to the aicore cluster.

Restart the Envoy and rerun the script (notice the new BASE_URL):

$ export ANTHROPIC_BEDROCK_BASE_URL=http://localhost:8100/bedrock
$ uv run -w anthropic["bedrock"] test-anthropic.py

[ToolUseBlock(id='toolu_bdrk_01G9A1A62B7a4RPGEUzVvHEZ', input={'command': 'ls -la *.py'}, name='bash', type='tool_use')]

Success! 🎉

Here is the full template: envoy-aicore.yaml

Connecting Claude Code

Anthropic’s code-generation tooling is extremely handy, so let’s wire it up.

Install Claude Code and configure additional settings:

export ANTHROPIC_BEDROCK_BASE_URL=http://localhost:8100/bedrock
export CLAUDE_CODE_SKIP_BEDROCK_AUTH=1
export CLAUDE_CODE_USE_BEDROCK=1
export ANTHROPIC_DEFAULT_SONNET_MODEL=d7ed726e32394096
export ANTHROPIC_DEFAULT_HAIKU_MODEL=d99de4372f34321e
export ANTHROPIC_MODEL=$ANTHROPIC_DEFAULT_SONNET_MODEL

Start the claude CLI and try something fun at the prompt:

> create a minesweeper game using js, html and css 

Claude Code will propose the file structure and generate the HTML, CSS and JavaScript automatically:

● I'll create a Minesweeper game for you with HTML, CSS, and JavaScript. Let me create the necessary files.

● Write(minesweeper.html)
  ⎿  Wrote 38 lines to minesweeper.html
...  

Enjoy your freshly generated Minesweeper game!

Recap

1. We routed Anthropic Bedrock traffic through Envoy, which injects the correct JWT and rewrites paths for SAP AI Core.
2. We successfully called the Sonnet-4.5 model from the Anthropic Python SDK.
3. We configured Claude Code to leverage the same deployment and generated a working game.

Your Envoy Proxy is now ready for further experiments — try it with Agents, MCP, or any other Bedrock-compatible tools!