Àkàndé

A self-hosted, provider-agnostic voice assistant that delivers structured executive briefings from any of ten LLM providers — including fully private local inference via Ollama and LM Studio.

---

Contents

Getting started

• Install — PyPI, source, system dependencies
• Quick Start — voice briefing in ten lines
• Use as a library — programmatic API

Surface

• Provider configuration — ten providers behind one env var
• Profiles and modes — AKANDE_PROFILE and AKANDE_MODE
• Interaction modes — TUI, classic CLI, web server, MCP
• Subcommands — data, verify-audit, mcp, install-local, skill
• Skills — briefing, web search, weather, finance + consent policy
• Model Context Protocol — serve and consume MCP

Operational

• Compliance — EU AI Act Article 50 controls
• Troubleshooting
• Trust — test count, coverage, security gates
• Development
• License

---

Install

Prerequisites

Dependency	Why	Ubuntu / Debian	macOS
Python 3.10+	Runtime	sudo apt install python3.12 python3.12-venv	brew install python@3.12
portaudio (optional)	Microphone capture ([mic] extra)	sudo apt install portaudio19-dev	brew install portaudio
ffmpeg	Audio decoding	sudo apt install ffmpeg	brew install ffmpeg

From PyPI

# Core install — provider SDKs and mic capture are optional extras.
pip install akande

# Full kit: every provider + microphone + MCP.
pip install "akande[all,mic,mcp]"

From source

git clone https://github.com/sebastienrousseau/akande
cd akande
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"        # bundles every test-relevant extra

Extras index

Extra	Pulls in	Enables
mic	pyaudio	Microphone capture (requires PortAudio system headers)
anthropic / google / mistral / cohere / huggingface / groq	the matching SDK	The corresponding LLM_PROVIDER value
offline-tts	pyttsx4	Offline TTS fallback
tts-local	kokoro-onnx	Local Kokoro-82M TTS (AKANDE_TTS=kokoro)
watermark	audioseal, torch	AudioSeal voice watermarking (Article 50 §2)
redact	presidio-analyzer	Higher-recall PII redaction in the cache
memory	mem0ai	Long-term memory façade
mcp	mcp	Run / consume Model Context Protocol servers
redis	redis	Distributed rate limiter for the web server
all	every provider SDK + pyttsx4	Provider-agnostic deployments
dev	testing + lint + audit + every provider SDK + mcp	Local development

---

Quick Start

# 1. Install the core + the mic extra.
pip install "akande[mic]"

# 2. Point at a provider.
export LLM_PROVIDER=openai
export OPENAI_API_KEY=sk-your-key-here

# 3. Launch the TUI.
akande

The TUI accepts spoken or typed questions and renders the briefing as it streams in. PDF and CSV artefacts are written to a date-keyed output directory on every answered question.

Use as a library

"""Ask Àkàndé a question programmatically."""

import asyncio

from akande.akande import Akande
from akande.providers import get_provider


async def main() -> None:
    # 1. Pick any of the ten configured providers by name.
    provider = get_provider("openai")  # honours $OPENAI_API_KEY
    akande = Akande(openai_service=provider)

    # 2. Ask a question.  The four-section briefing comes back as plain text.
    question = "What is quantitative easing?"
    response = await akande.openai_service.generate_response(
        user_prompt=question,
        system_prompt="You are an executive briefing assistant.",
        model="gpt-4o-mini",
    )

    # 3. Print the structured briefing.  `choices[0].message.content` follows
    # the OpenAI-shaped response envelope used by every provider.
    print(response.choices[0].message.content)


if __name__ == "__main__":
    asyncio.run(main())

---

Provider configuration

Set LLM_PROVIDER in your environment (or .env file). Each provider reads its own credentials from environment variables.

Provider	LLM_PROVIDER	Required env vars	Install	Default model
OpenAI	openai	OPENAI_API_KEY	(included)	gpt-3.5-turbo¹
Anthropic	anthropic	ANTHROPIC_API_KEY	pip install akande[anthropic]	claude-3-haiku-20240307
Google Gemini	google	GOOGLE_API_KEY	pip install akande[google]	gemini-pro
Mistral	mistral	MISTRAL_API_KEY	pip install akande[mistral]	mistral-small-latest
Cohere	cohere	COHERE_API_KEY	pip install akande[cohere]	command-r
Hugging Face	huggingface	HUGGINGFACE_API_KEY	pip install akande[huggingface]	mistralai/Mistral-7B-Instruct-v0.2
Groq	groq	GROQ_API_KEY	pip install akande[groq]	llama3-8b-8192
Azure OpenAI	azure_openai	AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT	(included)	gpt-35-turbo
Ollama	ollama	OLLAMA_HOST (optional)	(included)	llama3
LM Studio	lmstudio	LMSTUDIO_HOST (optional)	(included)	local-model

¹ Override per-call with the model argument or globally with OPENAI_DEFAULT_MODEL.

Install every provider SDK at once with pip install akande[all].

---

Profiles and modes

Àkàndé exposes two orthogonal sovereignty switches.

AKANDE_PROFILE selects the compliance posture:

Profile	EU residency	Audio watermark	Audit signing	Telemetry opt-in
local (default)	—	off	off	off
eu	enforced	on	on	off
strict	enforced	on	on	off
internal	—	on	on	on (opt-in only)

AKANDE_MODE selects the network posture:

Mode	Provider gate	Cache writes
online (default)	any provider	enabled
offline	ollama or lmstudio only	enabled

# EU-residency-aware cloud setup
export AKANDE_PROFILE=eu AKANDE_MODE=online

# Fully air-gapped local stack
export AKANDE_PROFILE=strict AKANDE_MODE=offline LLM_PROVIDER=ollama

---

Interaction modes

Mode	How to launch	What you get
TUI (default)	akande	Textual chat UI with streaming, voice toggle, history, export
Classic CLI	akande --classic	Numbered menu: voice, text, server, quit
Web server	akande → start server (or library Akande.start_server())	CherryPy server at http://127.0.0.1:8080 with SSE briefing endpoint
MCP server	akande mcp serve	Expose Àkàndé as MCP tools (Claude Desktop, Cursor, Continue)
Library	from akande.akande import Akande	Programmatic embedding

---

Subcommands

akande --help                  # top-level help
akande --version               # installed version

# GDPR data subject controls (export / delete)
akande data export --user alice --output alice.json
akande data delete --user alice --yes

# Audit verification — Ed25519-signed briefing sidecars
akande verify-audit  path/to/briefing.audit.json
akande verify-pdf    path/to/briefing.pdf
akande verify-watermark path/to/briefing.mp3 --threshold 0.5

# Model Context Protocol
akande mcp serve                # stdio MCP server (Claude Desktop ready)
akande mcp serve --http         # streamable HTTP transport
akande mcp list                 # list configured upstream servers
akande mcp list <server>        # introspect a server's tools

# One-shot fully-offline bootstrap
akande install-local --model llama3.1 --env-path .env

# Skill management
akande skill list
akande skill enable web_search
akande skill consent web_search
akande skill revoke web_search

---

Skills

Skills are specialised handlers the router picks over a generic LLM call. Five ship in the box; third-party skills register via the akande.skills entry-point group.

Skill	Match	Consent required	Offline-safe
briefing	default	no	yes
web_search	search, look up …, find …	yes	no
weather	weather in …, forecast …	no	no
finance	price of …, ticker …	no	no
policy (gate)	always — enforces consent	n/a	yes

"""Register a third-party skill via the entry-point group."""
# pyproject.toml
# [project.entry-points."akande.skills"]
# my_skill = "my_package.skill:MySkill"

from akande.skills.base import (
    Skill,
    SkillMeta,
    Intent,
    SkillContext,
    SkillResult,
)


class MySkill(Skill):
    @property
    def meta(self) -> SkillMeta:
        return SkillMeta(
            name="my_skill",
            description="One-line description of what this skill does.",
            requires_consent=True,
        )

    def match(self, text: str) -> Intent | None:
        if text.lower().startswith("my-skill:"):
            return Intent(name="my_skill", raw_text=text)
        return None

    def handle(self, intent: Intent, ctx: SkillContext) -> SkillResult:
        return SkillResult(content=f"Handled: {intent.raw_text}")

---

Model Context Protocol

Àkàndé can serve and consume MCP. The server exposes the briefing, audit, and skill surface as MCP tools; the client introspects upstream servers configured in ~/.akande/mcp.json.

# Serve over stdio for Claude Desktop / Cursor / Continue.
akande mcp serve

# Or streamable HTTP for HTTP-only hosts.
akande mcp serve --http

Claude Desktop drop-in (claude_desktop_config.json):

{
  "mcpServers": {
    "akande": {
      "command": "akande",
      "args": ["mcp", "serve"]
    }
  }
}

---

Compliance

Àkàndé ships the controls required by EU AI Act Article 50 (in force 2026-08-02) out of the box when AKANDE_PROFILE=eu (or strict):

• AI disclosure — every briefing carries a machine-readable disclosure block (akande.disclosure)
• AudioSeal watermark — synthesised audio is watermarked when the [watermark] extra is installed; absence is logged but never blocks
• Ed25519-signed audit sidecars — every PDF + CSV is paired with a .audit.json signed at write time; akande verify-audit re-verifies
• GDPR data export / delete — akande data export|delete against the SQLite conversation store
• Consent log — voice-cloning prompts require explicit consent recorded in the audit chain

---

Troubleshooting

Problem	Cause	Fix
Could not find PyAudio	PortAudio system headers missing	Ubuntu: sudo apt install portaudio19-dev. macOS: brew install portaudio. Then pip install akande[mic].
ffmpeg not found	ffmpeg not installed	Ubuntu: sudo apt install ffmpeg. macOS: brew install ffmpeg.
Microphone not detected	OS permissions	Grant microphone access in system settings.
ModuleNotFoundError: No module named 'anthropic'	Provider SDK not installed	pip install akande[anthropic] (or the relevant provider extra).
Invalid or missing OPENAI_API_KEY	Key not set or malformed	Ensure your environment or .env contains a valid sk- prefixed key.
AKANDE_MODE=offline forbids provider openai	Offline mode allows only local providers	Set LLM_PROVIDER=ollama or LLM_PROVIDER=lmstudio, or switch back to AKANDE_MODE=online.

---

Trust

• 785 tests + 95 % line coverage in CI on every push and pull request, on Python 3.10 / 3.11 / 3.12
• Quality gates: ruff (lint + format), mypy (strict islands on the provider surface), bandit (SAST), pip-audit (vulnerable-deps scan) — all blocking
• Fresh-install regression matrix (Ubuntu × 3.10/3.11/3.12 + macOS × 3.12) reproduces the user install path on every push
• Security posture documented in SECURITY.md: CSP nonces, custom-header CSRF, per-IP rate limiting (in-memory or Redis), CSV-formula injection prevention, filename sanitisation, IP hashing in logs

---

Development

git clone https://github.com/sebastienrousseau/akande
cd akande
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Quality gates (mirror CI)
ruff check . && ruff format --check .
mypy akande
pytest -q                # uses the [pytest] cov gate (95 %)
bandit -r akande
pip-audit

# Fresh-install regression on this machine
./scripts/regression.sh

See CONTRIBUTING.md for the full development loop.

---

License

Dual-licensed under the Apache License, Version 2.0 and the MIT License. See LICENSE-APACHE and LICENSE-MIT for the full text. You may pick whichever fits your project.