idun-sdk 0.1.12

Thin client + CLI for Azure AI Foundry agent NatureLM-Idun-5-MoE

Idun SDK

Thin, stdlib-only client + CLI for the NatureLM-Idun-5-MoE agent on Azure AI Foundry (codename Idun). No httpx, no azure-identity — it runs headless on Termux/Android with nothing but the Python standard library.

Idun is a tool agent: it reasons and calls tools (web_search, memory_search). This SDK surfaces the full agent trajectory — every reasoning step and tool call — instead of a black-box chatbot wheel, so you can see how it arrived at an answer.

Install

pip install idun-sdk

The idun CLI and the idun Python package (plus the stdlib MCP server idun_mcp.py) are installed.

Authenticate (device-code, Entra)

idun login
# opens https://microsoft.com/devicelogin — enter the printed code,
# sign in with your tenant account (no admin role required).
# Token is saved to ~/foundry_token.txt (FOUNDRY_TOKEN).

No admin role needed. idun login runs a standard Entra device-code flow and returns a plain bearer access token for the Foundry agent endpoint. You do not become a portal admin — you only get an API token that the SDK attaches to each request. Any tenant user granted access to the NatureLM-Idun-5-MoE agent (RBAC on the Foundry resource) can run idun login. Admin rights are only needed to configure or redeploy the agent, not to use it.

The token is a normal Entra FOUNDRY_TOKEN (scope https://ai.azure.com/.default); it auto-rotates before expiry (Phase 2.5).

Use

# final answer only
idun chat "Fasse in einem Satz zusammen, was Contoso im Bereich Nachhaltigkeit kommuniziert."

# full agent trajectory (reasoning + web_search tool steps)
idun trace "Use web_search to find the current CEO of Contoso and report the name."

Python

from idun import IdunClient

res = IdunClient().complete("Your prompt here")
print(res.text)            # final answer
for s in res.steps:        # agent trajectory
    if s.kind == "tool":
        print("TOOL", s.tool, s.status, s.query)
    else:
        print("REASON", s.text[:80])

IdunResult has:

• text — the final answer
• steps — list of Step (kind="reasoning"|"tool", plus tool/query/status for tool steps)
• model — the backing model (e.g. gpt-5.4-2026-03-05)
• raw — the verbatim Foundry responses payload

Request shape (verified working)

POST {base}/api/projects/{project}/agents/{agent}/endpoint/protocols/openai/responses?api-version=2025-05-15-preview
Authorization: Bearer ***
Content-Type: application/json

{"model": "model-router", "input": "<prompt string>", "max_output_tokens": 4096}

Notes:

• model MUST be "model-router" (the agent id is already in the URL).
• Do not send a tools key — the agent owns its capabilities; doing so returns 400 invalid_payload.
• The answer is in output[].content[].text; tool calls appear as web_search_call items with action.queries and status.

MCP — agent + docs

Idun is available as an MCP server and has a GitMCP docs mirror, so other agents can both call Idun and read its documentation without hallucinating.

1. Idun MCP server (stdlib-only, local)

idun_mcp.py is a zero-dependency stdio MCP server — no FastMCP / httpx needed (runs on bare Python, ideal for Termux/Android).

python3 idun_mcp.py        # stdio MCP server

Tools exposed:

• idun_chat(prompt) — final answer text
• idun_trace(prompt) — full agent trajectory (steps + text)

Add to any MCP client (e.g. Cursor ~/.cursor/mcp.json):

{
  "mcpServers": {
    "idun": { "command": "python3", "args": ["/abs/path/idun-sdk/idun_mcp.py"] }
  }
}

2. GitMCP docs mirror (remote, zero-setup)

https://gitmcp.io/qapdex-maker/idun-sdk/sse

For stdio-only clients (Claude Desktop, Cline, Msty):

{ "mcpServers": { "idun-docs": { "command": "npx", "args": ["mcp-remote", "https://gitmcp.io/qapdex-maker/idun-sdk/sse"] } } }

Recommended combo for a foreign agent: both idun (calls the agent) and idun-docs (reads the SDK docs) — it can invoke Idun and look up the exact IdunClient signature on its own.

Why stdlib-only?

Many AI agents run on constrained hosts (Termux on Android, minimal containers). Pulling httpx/pydantic/azure-identity breaks those setups. This SDK uses urllib.request + json only, so pip install idun-sdk just works everywhere Python 3.8+ runs.

Links

• PyPI: https://pypi.org/project/idun-sdk/
• SDK repo: https://github.com/qapdex-maker/idun-sdk
• Playground repo: https://github.com/qapdex-maker/idun-playground
• GitMCP docs: https://gitmcp.io/qapdex-maker/idun-sdk/sse