armance 0.2b3

Armance 👒 - a team of experts you recruit to help you decide. Coordinate a panel that debates, stress-tests and synthesises. You keep the call 💡. Not a copilot, a brain you consult when the choice matters.

Armance 👒

A brain you consult when the choice matters.

Not a copilot. Not an autonomous agent. A small firm of LLM experts that debates, stress-tests, and synthesises — over your own documents — so you can make a decision you can defend.

Install · Quickstart · How it works · Contributing · Vision · Architecture

---

For AI agents reading this file: start here, then read roadmap/02_architecture.md for the module map and CLAUDE.md for project conventions.

---

Why Armance

Most "AI tools" are makers — they race to a single answer, then ask you to ship it. That is the wrong shape for the hardest moments of strategic work:

• Framing an ambiguous problem.
• Deciding between two plausible directions.
• Stress-testing a plan before committing.
• Synthesising contradictory expert opinions.

In those moments you do not need a faster typist. You need a thinking partner that pulls multiple competent perspectives, makes them disagree productively, and forces you to see the angle you would have missed.

Armance is that partner. It runs locally, reads your documents, recruits a panel of specialists with conflicting personas, runs them through a workflow, and hands you a defensible brief. You keep the call.

The brain-vs-hands principle

Armance does	Armance does not
Frame the problem with you	Write your codebase
Recruit a panel of disagreeing personas	Manage a queue of tasks
Run a workflow across that panel	Push to your CI
Stress-test the synthesis	Send the email
Produce a defensible decision brief	Be your IDE

---

Meet the staff

Five permanent meta-agents live in every Armance project. They never recruit themselves; they coordinate the specialists Armance brings in for your specific brief.

	Role	Voice
Armance	Host — frames the project, routes the room	French refinement. Systematic vouvoiement.
Malik	Recruiter — picks specialists whose personas disagree usefully	Quiet force, slow tempo, Gainsbourgian nonchalance.
Kim	Operator — designs and runs the workflow	Direct, technical, square. Numbered lists, no padding.
Mona	VP — synthesises, challenges, recommends	Cash and direct. Changes a room's centre of gravity in two sentences.
Serge	Adversarial criticalist — stress-tests every synthesis	The cynical senior French engineer. « *soupir* » before every critique.

Plus the specialists Malik recruits per project (UX researchers, historians, ops engineers — whatever the brief calls for), each with a generated 200-word persona prompt.

---

Install

Project home: armance.io (site coming soon). For the full, OS-by-OS guide (and a copy-paste prompt to let an AI agent install it for you), see INSTALL.md.

Prerequisites

• Python ≥ 3.11
• At least one LLM provider key. OpenRouter is the easiest start — its free tier runs Armance end to end with no spend.

The one command

pip install armance

That's it. The base install gives you the full CLI/TUI, RAG, all four providers (OpenRouter, claude-code subscription, Gemini, custom OpenAI-compatible), DOCX / PPTX / Markdown deliverables, and the web UI — no extras to choose. (PDF export is the one optional extra — see below — because it needs native libraries pip can't install on its own.)

Prefer an isolated tool install? uv tool install armance or pipx install armance work identically.

Installing from git (pip install git+https://github.com/armance-io/armance.git) ships the CLI but not the prebuilt web UI — the bundle is a build artifact, not tracked in git. From a git install, run armance web --build once (needs Node + pnpm) to generate it, or use a release wheel from PyPI where the UI is already bundled.

Verify:

armance --version
armance doctor      # config, provider reachability, sqlite-vec, deliverables, ledger

Reaching Claude models: two ways, both on the base install. Through OpenRouter (set OPENROUTER_API_KEY), or through a Claude Pro/Max subscription via the bundled claude-code provider (Anthropic's claude-agent-sdk ships in the package — no extra to install).

Optional extra: PDF export

Every export format works out of the box except PDF. PDF uses WeasyPrint, which depends on native libraries (GTK / Pango / Cairo) that pip cannot install — so it is opt-in:

pip install 'armance[pdf]'

Then install the native libs for your OS:

• Linux: sudo apt-get install libgobject-2.0-0 libcairo2 libpango-1.0-0
• macOS: brew install pango (usually already present)
• Windows: install the GTK runtime — see the WeasyPrint install guide.

DOCX / PPTX / Markdown exports need none of this. If WeasyPrint is missing, armance simply skips PDF with a clear message — it never crashes.

Contributor / dev setup

git clone https://github.com/armance-io/armance.git
cd armance
uv sync

An editable install lives in the project's .venv — it is not put on your PATH. Run the CLI with uv run armance … (or activate the venv: source .venv/bin/activate). A separate uv tool install / pipx install is what puts a global armance on PATH; don't confuse the two.

The web UI is a build artifact and is not in git, so a fresh clone has no bundle yet. Build it once (needs Node + pnpm), then run:

uv run armance web --build          # builds the UI bundle, then serves it
# subsequent runs (bundle already built):
uv run armance web

See web/DEVELOPMENT.md for the full dev loop (hot-reload vs bundled, tests).

---

Running the web client (UI)

Armance V2 ships a Belle Époque web UI (Next.js 16 + React 19) on top of a FastAPI server. One command runs both the API and the UI in a single process — no Node, no second server:

armance web                 # serves API + UI at http://127.0.0.1:8000, opens a browser

By default the server runs in the background: the command waits until the server answers, opens a browser, prints the URL + pid, and returns 0. Logs go to .armance/logs/web-server.log (not the terminal). Stop it with armance web --stop. Use --foreground to block instead (Ctrl+C stops it; logs stream to the terminal) — handy for dev.

Only one instance per project folder may run at a time: a lock file at .armance/web-server.pid records the running server. A second armance web in the same folder is refused until you stop the first (stale locks from a crashed server are detected and cleared automatically).

Run it from a project directory (one that has — or will have — a .armance/ folder, exactly like armance run). Options:

Flag	Effect
--port 8000	port (default 8000)
--bind 0.0.0.0	expose on the LAN (read-only for watchers; only the first client may write)
--no-browser	don't auto-open a browser
--foreground	block in the terminal (Ctrl+C stops; logs stream to stdout) instead of backgrounding
--stop	stop the server running in this folder, then exit (alias: armance web stop)
--build	(repo checkout only) rebuild the UI bundle before serving — needs Node + pnpm

Where the UI comes from:

Install method	Ships the UI?	How to get the UI
pip install armance (PyPI release)	✅ bundled	nothing — armance web just works
git clone + editable / uv tool install git+…	❌ artifact, not in git	run armance web --build once from a checkout (the tool install has no frontend sources — clone the repo)

If you see “no bundled UI found — running API only”, the bundle hasn't been built for this install. Use a release wheel, or armance web --build from a clone.

Frontend dev mode (hot reload)

For live iteration on the UI, run the two dev servers side by side (no bundle needed):

uv run armance web --no-browser              # API on :8000
cd web/frontend && pnpm install && pnpm dev  # UI on :3000, proxies /api → :8000

---

Quickstart

mkdir my-project && cd my-project
armance init        # interactive — providers, default model, budget, language
armance run         # opens the TUI

armance init walks through:

1. Providers — openrouter, claude-code, gemini, custom-openai (multi-select).
2. API keys — stored in .armance/.env (gitignored).
3. Default provider + model.
4. Budget effort — free-first / low / medium / high / adaptive.
5. Interface language — EN / FR / ES / DE / ZH / JA. Every agent — staff and specialists — replies in this language. Auto-detected from $LANG.
6. Embedding model — discovered from your configured provider's API; can be skipped and configured later.

Drop documents into .armance/docs/ and run armance run. Armance greets you, frames the project, recruits a team, designs a workflow, runs it, and exports a deliverable.

---

How it works

The user journey

                 ┌─────────────────┐
              1. │  armance init   │  pick providers, models, budget, language
                 └────────┬────────┘
                          │
                 ┌────────▼────────┐
              2. │   drop docs     │  .armance/docs/  → auto-indexed on `armance run`
                 └────────┬────────┘
                          │
                 ┌────────▼────────┐
              3. │   armance run   │  TUI opens; Armance greets in your language
                 └────────┬────────┘
                          │
                 ┌────────▼─────────────────────────────────────┐
              4. │  Armance frames the project                  │  asks focused questions,
                 │    "What audience? What constraint?"         │  proposes /library index
                 │    /library load <file> if needed            │  and /library load,
                 │  → /save when context is rich                │  freezes L0 context
                 └────────┬─────────────────────────────────────┘
                          │
                 ┌────────▼─────────────────────────────────────┐
              5. │  Malik recruits specialists                  │  axis of disagreement
                 │    "Sarah · data-driven / Julian · empathic" │  per role; rich
                 │    each with a 200-word persona              │  persona generation
                 └────────┬─────────────────────────────────────┘
                          │
                 ┌────────▼─────────────────────────────────────┐
              6. │  Kim designs a workflow                      │  3 strategies —
                 │    NL dialogue, narrow scope, tailored steps │  rapide / équilibrée /
                 │    interactive ↔ autonomous run mode         │  approfondie
                 └────────┬─────────────────────────────────────┘
                          │
                 ┌────────▼─────────────────────────────────────┐
              7. │  /workflow run <name>                        │  pre-run health-check,
                 │    parallel deliberation per DAG level       │  HITL checkpoints
                 │    Mona synthesises, Serge red-teams         │  (autonomous = Mona
                 │                                              │   proxies for the CEO),
                 │                                              │  versioned manifest
                 │                                              │  with per-step tokens
                 │                                              │  and duration
                 └────────┬─────────────────────────────────────┘
                          │
                 ┌────────▼─────────────────────────────────────┐
              8. │  /deliverable pdf | docx | pptx | md         │  .armance/exports/
                 └──────────────────────────────────────────────┘

Storage layout — everything is a file

.armance/
  docs/               your documents (PDF, DOCX, MD, TXT)
  vector/             sqlite-vec store + manifest.json + read.json
  agents/             one .md per agent (YAML frontmatter + system prompt)
    system-*.md         built-in staff (Armance / Malik / Kim / Mona / Serge)
    <Name>.md           Malik-recruited specialists, with rich personas
  workflows/          *.yaml DAG definitions
  context/            L0_v<N>.md / L1_<role>_v<N>.md / L2_<theme>_v<N>.md
  reports/            versioned <agent>_v<N>.md per step
  sessions/<id>/      state.json + ledger.json + conversation.md
  exports/<wf>/run-<ts>/  per-run manifest + per-step outputs + synthesis
  config.yaml         non-secret config (providers, default model, language)
  .env                provider API keys — gitignored

Markdown is the source of truth. SQLite is used only for vector retrieval.

Four-layer architecture

┌────────────────────────────────────────────────────┐
│ client     TUI (Textual) — and future web client   │
├────────────────────────────────────────────────────┤
│ transport  DTOs + event bus — wire-format only     │
├────────────────────────────────────────────────────┤
│ service    orchestration, agents, workflows, RAG   │
├────────────────────────────────────────────────────┤
│ core       pure models + protocols, no I/O         │
└────────────────────────────────────────────────────┘

Each layer imports only from layers below. Enforced by import-linter. See roadmap/02_architecture.md for the module map.

Providers

Provider	Key env var	Notes
openrouter	OPENROUTER_API_KEY	Default. Live model discovery. Many :free models. Supports reasoning effort.
claude-code	uses claude-agent-sdk auth	Bundled by default (no extra). Subscription = effectively free for the user. Native web search via WebSearch tool.
gemini	GEMINI_API_KEY	Live discovery via /v1beta/models. Native Google Search grounding.
custom-openai	CUSTOM_OPENAI_API_KEY + CUSTOM_OPENAI_BASE_URL	Any OpenAI-compatible endpoint.

Models are discovered live; pricing tiers, web-search capability, and reasoning support are read from each provider's API. No hardcoded catalogue.

Configuration

All non-secret settings live in .armance/config.yaml. API keys live in .armance/.env. Both are created by armance init and can be edited by hand — changes take effect on the next armance run.

.armance/config.yaml

Field	Default	What it does
default_provider	openrouter	LLM provider used for all agents unless overridden per-agent.
default_model	(chosen at init)	Model id for all agents. Per-agent override via the agent's .md frontmatter.
budget_effort	free-first	Cost guard. free-first keeps effectively-free models (OpenRouter :free, Claude subscription) regardless of nominal tier. Changeable live via /effort.
budget_cap_usd	null	Hard USD cap per session. null = no cap.
language	(chosen at init)	en / fr / es / de / zh / ja. All agents reply in this language.
embedding_provider	(chosen at init)	Provider used for RAG. Leave blank to disable.
embedding_model	(chosen at init)	Embedding model id. Must match embedding_provider.
log_level	INFO	DEBUG, INFO, WARNING, ERROR.
prices	{}	Override per-model USD prices. Example: prices: {my-model: {input_per_mtok: 1.0, output_per_mtok: 5.0}}.

Le modèle de Mona = default_model ; ajustez-le par agent via Malik si besoin.

.armance/.env

Variable	Provider	Notes
OPENROUTER_API_KEY	openrouter	Get one at openrouter.ai.
OPENROUTER_BASE_URL	openrouter	Override base URL (default: https://openrouter.ai/api/v1).
GEMINI_API_KEY	gemini	Google AI Studio.
GEMINI_BASE_URL	gemini	Override base URL.
ANTHROPIC_API_KEY	claude-code	Used for live model discovery via /v1/models.
CUSTOM_OPENAI_API_KEY	custom-openai	Custom endpoint API key.
CUSTOM_OPENAI_BASE_URL	custom-openai	e.g. http://localhost:11434/v1 for Ollama.

Shell values override .env values.

TUI commands (short list)

Command	Effect
/help, /quit	self-explanatory
/switch <agent>	route next turn to an agent (or @Name inline)
/save	freeze current project context into L0
/workflow design <name>	start Kim's workflow design dialogue
/workflow run <name>	execute a workflow (interactive or autonomous)
/deliverable pdf|docx|pptx|md	export the latest synthesis
/report	persist the last reply as a versioned report
/export claude|opencode|cline|roo|all	emit agent docs for another tool
/model, /effort	switch provider/model or reasoning effort

Everything is natural-language first. Slash commands are aliases. "Malik, recrute deux historiens" — the recruiter intercepts it.

What runs your turn

1. dispatch_input (service/tui_bridge) — routes by @mention or current agent.
2. The right meta-agent chat handler builds its system prompt: persona + voice overlay (your language) + RAG injection + project brief + team roster.
3. call_with_ledger calls the provider via the chosen LLMClient and accumulates token usage.
4. The reply is scanned for [EXECUTE:/save], [EXECUTE:/recruit], [EXECUTE:/workflow-design], [EXECUTE:/workflow-run:<name>:<mode>], [EXECUTE:/dismiss-all]. Tags trigger the corresponding side effect.
5. The conversation is appended to .armance/sessions/<id>/conversation.md and the ledger is persisted.

---

Tests

uv run pytest tests/                                       # core: unit + integration (no network)
bash scripts/check_invariants.sh                           # layer + lifecycle invariants
uv run python scripts/qa_live.py                           # live OpenRouter free-model QA

# Web backend (ships in the package):
cd web && uv run pytest ../src/armance/web/backend/tests/

# Web frontend:
cd web/frontend && pnpm typecheck && pnpm lint && pnpm test && pnpm e2e

qa_live.py exercises the full user journey: greeting → context → recruit → dismiss → re-recruit → design → run → deliverable → RAG round-trip → language switch.

---

Roadmap & vision

See roadmap/:

• 01_vision.md — the brain-vs-hands manifesto.
• 02_architecture.md — module map & dependency rules.

Phased planning and per-issue specs are tracked privately by the maintainer.

---

License

AGPL-3.0-or-later. Copyright © 2026 Guillaume Richard.

The strong copyleft license is intentional. If you run a modified Armance as a network service, you must publish your modifications. Commercial dual-licensing is available — contact [e-mail].

See LICENSE for the full text.