AegisAI
Setup

Three commands. One file.

1. Install

pip install anthropic httpx

That's it. The Anthropic SDK and an HTTP client.

2. Set env vars

export ANTHROPIC_API_KEY=sk-ant-...
export AEGISAI_BASE_URL=https://api.aegisai.store
export AEGISAI_USER_TOKEN=<user JWT>

3. Run the example

python example.py "Show me vendor 100247's open invoices"

Claude reasons, emits a tool_use block, the dispatch loop hits AegisAI, the masked response goes back to Claude.

What's in the kit

Three files in samples/connectors/anthropic-claude/.

FilePurpose
README.mdSetup walkthrough plus the two-token identity model explainer.
tool_spec.jsonThe tool definition you pass to the Anthropic SDK (or Bedrock/Vertex equivalent).
example.pyWorking ~80-line example with the full dispatch loop.
The two-token rule

Anthropic's key authorises you. The user's JWT authorises the user.

Two tokens, never confused: the Anthropic API key proves your application is allowed to call Claude. The end user's JWT proves the user is allowed to access the SAP/cloud data. Your application is the bridge — Claude generates a tool_use block, your dispatch loop attaches the user's JWT in the request body, AegisAI extracts sub, and the audit row records the human.

If you collapse these into one token (e.g. mint a fake user identity from the Anthropic key), the audit chain shows "the bot" for every action and identity propagation collapses. Don't.

Building a Claude-powered ops tool?

If you're standing up an internal Claude agent that talks to SAP, we can pair on the dispatch loop, the audit-correlation IDs, and the per-tool denial UX.

Email us