Idun SDK

Thin, stdlib-only client + CLI for the NatureLM-Idun-5-MoE agent on Azure AI Foundry — with a pluggable multi-backend layer.

Runs headless on Termux/Android with nothing but the Python standard library (no httpx, no azure-identity, no Flask). Idun is a tool agent (reasons + calls tools like web_search); this SDK surfaces the full agent trajectory (reasoning + tool calls) instead of a black-box answer.

Multi-backend

The same IdunClient API dispatches to four interchangeable backends. Non-Azure backends need no FOUNDRY_TOKEN, so the SDK keeps working even when the Azure subscription is suspended.

Set the backend globally via env (IDUN_BACKEND=hf) or per-call (--backend). Model overrides: HF_MODEL, GITHUB_MODEL, OLLAMA_BASE, OLLAMA_MODEL. Non-Azure backends return a flat answer (res.steps == []) — the tool-agent trajectory is an Azure-Idun feature.

Install

pip install idun-sdk

Installs the idun CLI, the idun Python package, and the stdlib MCP server idun_mcp.py. Requires Python ≥ 3.8; no third-party dependencies (stdlib-only, runs headless on Termux/Android).

Verify the install:

idun --help          # shows all commands
idun welcome         # banner + matrix intro

Setup

Idun is backend-agnostic. Pick a backend once with the wizard (writes ~/.idunrc, sourced automatically by your shell), or set credentials per backend. Run the wizard for a guided, universal first-run setup:

idun wizard          # choose backend, capture creds/config -> ~/.idunrc
source ~/.idunrc     # or restart your shell
idun status          # confirm active backend + credential state

Backend credentials

Azure (default). idun login runs a device-code flow and stores an Entra bearer token. No admin role required — any tenant user with agent RBAC can run it; admin rights are only needed to configure the agent, not to use it. The token auto-rotates before expiry.

Hugging Face / GitHub. The wizard or idun login --backend <x> prompts for a token and saves it (0600). Both offer a free tier; no Azure subscription or card needed.

Ollama. Run a local Ollama server (e.g. ollama serve) and point Idun at it:

export OLLAMA_BASE=http://localhost:11434
export OLLAMA_MODEL=llama3.1

Switching backends

Globally via env, or per call via --backend:

export IDUN_BACKEND=hf          # all future calls use HF
idun chat --backend github "Hi" # one-off override

Model overrides (optional): HF_MODEL, GITHUB_MODEL, OLLAMA_BASE, OLLAMA_MODEL.

Quick test

idun chat "Hello"                       # azure (needs login first)
idun chat --backend hf "Hello"          # any backend with creds set

Quickstart

idun chat "Summarize Contoso's sustainability comms in one sentence."
idun trace --backend azure "Use web_search to find the current CEO of Contoso."

from idun import IdunClient

# Azure (default)
res = IdunClient().complete("Your prompt here")
print(res.text)                 # final answer
for s in res.steps:             # agent trajectory
    print(s.kind, s.tool, s.query, s.status)

# Hugging Face (no Azure token needed)
res = IdunClient(backend="hf", hf_token="hf_xxx",
                 hf_model="microsoft/phi-3-mini-4k-instruct").complete("Hello")
print(res.text)

Commands

Hugging Face pipeline

The hf command wraps the Hugging Face Hub + Inference API (stdlib-only for whoami / status; push uses the optional huggingface_hub client):

idun hf whoami                              # validate token, show HF user
idun hf status microsoft/phi-3-mini-4k-instruct   # exists? gated? private? task?
idun hf push Qapdex/my-agent-out a.txt b.txt       # create repo + upload files

push needs pip install huggingface_hub (kept optional so the SDK stays stdlib-only for everything else). Upload under your user namespace (Qapdex/...), not an org, unless your token has write rights there.

MCP server

idun_mcp.py is a zero-dependency stdio MCP server (no FastMCP / httpx):

python3 idun_mcp.py

Exposes idun_chat(prompt) and idun_trace(prompt). Add to any MCP client:

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

Docs mirror (GitMCP): https://gitmcp.io/qapdex-maker/idun-sdk/sse

Combine with Sentry (MCP)

Idun pairs well with Sentry's MCP server so an agent can both call Idun and inspect Sentry errors. Add it remotely (OAuth, note the /mcp suffix):

{ "mcpServers": { "sentry": { "url": "https://mcp.sentry.dev/mcp" } } }

Recommended combo: idun (calls the agent) + idun-docs (reads SDK docs) + sentry (queries error tracking).

Azure request shape (verified)

POST {base}/api/projects/{project}/agents/{agent}/endpoint/protocols/openai/responses?api-version=2025-05-15-preview
Authorization: Bearer ***
{"model": "model-router", "input": "<prompt>", "max_output_tokens": 4096}

• model MUST be "model-router" (agent id is in the URL).
• Do not send a tools key — the agent owns its tools (else 400 invalid_payload).

Links

• PyPI: https://pypi.org/project/idun-sdk/
• Repo: https://github.com/qapdex-maker/idun-sdk
• Foundry: https://ai.azure.com