A thin, dependency-free Python client + CLI for Azure AI Foundry agents (Idun / NatureLM) and any OpenAI-compatible endpoint. Tenant-agnostic by design: bring your own Foundry resource — no hardcoded tenant coordinates.
pip install idun-sdk
# with optional PDF ingest:
pip install "idun-sdk[pdf]" # pulls PyPDF2from idun import IdunClient
client = IdunClient() # reads ~/.idun/config.toml
out = client.complete("Summarize the quarterly risk report.")
print(out["text"])This package ships two command-line tools with distinct purposes. They are
intentionally kept separate — idun is the Azure AI Foundry client,
idun-multi is the multi-provider LLM console. There is no unified
idun-wizard; each tool has its own first-run setup wizard:
idun— the Azure AI Foundry client. Agent completions, trajectory export, document matrix (idun matrix), prompt packs (idun run/idun packs), and Hugging Face Hub operations (idun hf). Azure-specific.idun wizardconfigures the Azure Foundry client (endpoint / project / agent) in~/.idun/config.toml.
idun-multi— the multi-provider LLM console. Talks to any of the 17 registered providers (OpenAI, Anthropic, Groq, OpenRouter, HF, …) plusrace,cost,models,doctor,support. Provider-agnostic.idun-multi wizardconfigures the default LLM provider (picks one of the 17 registered providers and prints the provider table so you can choose) in~/.idun/config.toml.
Both wizards write only to ~/.idun/config.toml (never ~/.idunrc,
which is no longer used). Credentials live in per-provider ~/.idun/<id>.token
files (0600).
idun-multi providers # list providers + credential status
idun-multi -p openrouter ask "hello" # -p goes BEFORE the subcommand
idun-multi race "explain quantum" # compare providers on one prompt
idun-multi doctor # environment + credential audit (also checks scripts)
idun-multi wizard # pick the default LLM provider
idun wizard # configure the Azure Foundry client
import asyncio
from idun import AsyncIdunClient
async def main():
c = AsyncIdunClient()
out = await c.acomplete("What changed vs last quarter?")
print(out["text"])
asyncio.run(main())idun — Azure AI Foundry client:
idun chat "your prompt" # one-shot completion
idun trace "your prompt" # full agent trajectory (steps)
idun export "your prompt" -o trace.md # run + save trajectory (json/md)
idun token # inspect / rotate stored Entra token
idun login # device-code login to your Foundry (azure/hf/openai)
idun status # show resolved backend + credential state
idun run <pack> <key> # run a bundled prompt pack
idun packs # list bundled prompt packs
idun matrix --docs DIR --questions FILE # Doc x Question pivot (IDEA α)
idun diff-docs --doc-a A --doc-b B --topics T # clause-drift compare (IDEA γ)
idun diff "A" "B" # compare two prompt trajectories side-by-side
idun hf whoami|status|push # Hugging Face Hub operations
idun openapi # print the bundled OpenAPI 3 spec for the completion API
idun logo # show bundled Foundry logo paths
idun welcome # ASCII banner
idun wizard # configure the Azure Foundry client
idun-multi — multi-provider LLM console (-p <provider> goes BEFORE the subcommand):
idun-multi providers # list providers + credential status
idun-multi -p openrouter ask "hello" # chat with a provider
idun-multi race "explain quantum" # compare providers on one prompt
idun-multi models # list models (or --discover)
idun-multi cost # token / cost estimate
idun-multi doctor # environment + credential audit
idun-multi verify # LIVE smoke-test configured providers
idun-multi wizard # pick the default LLM provider
idun-multi review <pr> [--post] [--inline] [--labels] [--no-cache]
# self-built PR reviewer (diff -> race -> artifacts)
The support matrix is honest about capability (which transport is wired),
but not about whether a provider has actually answered a request lately.
idun-multi verify performs a real, minimal API call against every provider
that has a credential configured, and records the outcome in
~/.idun/.verified.json (no secrets — only status, model, latency, and a
redacted error on failure). Unconfigured providers are reported as skip,
never fail, so an unconfigured machine shows an honest "not checked"
rather than a wall of false failures.
The recorded state feeds the Live column of idun-multi support and is
refreshed automatically every time you run race or verify. A provider with
Declared = — and Live = ? is simply unproven on this install — not
claimed broken, not claimed working.
idun-multi verify # all providers with a credential
idun-multi verify --providers openai,groq,anthropic
A self-hosted alternative to CodeRabbit (ROADMAP Open #4). It fetches the PR
diff via gh, splits it into chunks, races them across up to 3 providers that
have credentials, and turns the results into actionable artifacts:
idun-multi review 123 # dry-run, prints findings
idun-multi review 123 --post # also post a PR comment
idun-multi review 123 --inline # post inline comments (GraphQL)
idun-multi review 123 --labels --post # add severity labels to the PR
idun-multi review 123 --no-cache # re-run all providers
Findings are parsed into severity/file/line and deduplicated across providers.
Chunk results are cached under ~/.idun/.review_cache.json (no secrets) so
repeat runs are cheap.
The idun CLI / IdunClient speak to any OpenAI-compatible endpoint plus the
native Azure AI Foundry transport. Registered providers (4 groups):
- OpenAI-compatible:
openai,groq,together,perplexity,fireworks,novita,xai,deepseek,mistral,openrouter,gemini,hf,nous - Anthropic:
anthropic(Claude) - Azure:
azure(Azure AI Foundry / Idun agent) - Local:
ollama,local(bring your own endpoint)
Each provider's capabilities (streaming, tools, vision, JSON mode) are derived
from the transports actually implemented in idun/providers.py — see
SUPPORT_MATRIX.md for the full, code-generated
matrix and per-transport capability details.
Hugging Face (hf): uses the OpenAI-compatible router
https://router.huggingface.co/v1 (the legacy api-inference.huggingface.co
host was retired). Requires an HF token (HF_TOKEN).
Any OpenAI-compatible base URL works with zero code changes — set it in the config and the provider switches automatically.
~/.idun/config.toml (neutral defaults — supply your own resource). Secrets go
to per-provider ~/.idun/<id>.token files (mode 0600), not in this file:
# Written by `idun-multi wizard` (LLM provider default):
[defaults]
provider = "openrouter"
[openrouter]
model = "deepseek/deepseek-chat"
# Written by `idun wizard` (Azure AI Foundry client):
[azure]
base = "https://YOUR-RESOURCE.services.ai.azure.com"
agent = "your-agent-name"
project = "your-project"No tenant coordinates are baked into the shipped code. The CLI reads the config
via idun.config; environment variables (IDUN_PROVIDER, OPENAI_API_KEY, …)
always win over the file.
These build on the core SDK and live in
idun.matrix/ the playground repo.
Build an N × M answer matrix over documents × questions. Each cell carries the answer, the source citation, and a status (GREEN = answered+cited, RED = contradiction, GRAY = no info).
idun matrix --docs ./contracts --questions ./questions.txtCompare two documents across topics and flag deviations:
idun diff-docs --doc-a contract_a.txt --doc-b contract_b.txt --topics topics.txtA tenant-agnostic mobile web UI that runs idun matrix against your Foundry
resource: https://qapdex-maker.github.io/idun-playground/matrix_app.html
- Live demo (recorded trajectories + matrix UI): https://qapdex-maker.github.io/idun-playground/
- Matrix concept note: https://github.com/qapdex-maker/idun-playground/blob/main/DOC_MATRIX_CONCEPT.md
- Source: https://github.com/qapdex-maker/idun-sdk
- Changelog: https://github.com/qapdex-maker/idun-sdk/blob/main/CHANGELOG.md
pip install -e ".[dev]"
pytest # offline; no network/API keys required
ruff check . # lint (pinned ruff==0.15.10 in CI)MIT.