tramontane 0.1.6

Mistral-native agent orchestration framework — EU sovereign, GDPR built-in

TRAMONTANE

Mistral-native agent orchestration. Not bolted on top.

---

Why Tramontane

• Mistral-native -- built FROM Mistral's primitives (Agents API, Conversations, Handoffs), not adapted from OpenAI patterns
• Smart router -- Ministral-3B classifies your task, then routes to the optimal model from a 10-model fleet. You set model="auto" and a EUR budget ceiling; the router handles the rest
• EU-sovereign -- GDPR built-in (PII detection, Article 30 reports, right to erasure), EUR cost tracking, Scaleway fr-par deployment

Quick Start

pip install tramontane
export MISTRAL_API_KEY=your_key_here

import asyncio
from tramontane.core.agent import Agent
from tramontane.router.router import MistralRouter

async def main():
    agent = Agent(
        role="Code Reviewer",
        goal="Review code for bugs and security issues",
        backstory="Senior engineer with 10 years experience",
        model="auto",       # router picks the optimal Mistral model
        budget_eur=0.01,    # hard cost ceiling in EUR
    )
    router = MistralRouter()
    result = await agent.run("def add(a, b): return a + b", router=router)

    print(f"Model: {result.model_used}")      # e.g. devstral-small
    print(f"Cost:  EUR {result.cost_eur:.4f}") # e.g. EUR 0.0003
    print(result.output)

asyncio.run(main())

Or run a multi-agent pipeline from YAML:

tramontane run pipelines/code_review.yaml --input "def add(a,b): return a+b" --budget 0.05

How the Router Works

Every prompt goes through a Ministral-3B classifier that determines task type and complexity, then the router picks the cheapest capable model:

has_vision?              -> pixtral-large
task=code, complexity<4  -> devstral-small    (EUR 0.10/0.30 per 1M)
task=code, complexity>=4 -> devstral-2        (EUR 0.50/1.50 per 1M)
needs_reasoning, <4      -> magistral-small   (EUR 0.50/1.50 per 1M)
needs_reasoning, >=4     -> magistral-medium  (EUR 2.00/5.00 per 1M)
task=bulk                -> ministral-7b      (EUR 0.10/0.10 per 1M)
default                  -> mistral-small     (EUR 0.10/0.30 per 1M)

When budget forces a downgrade, quality floors prevent garbage output:

Task Type	Floor Model	Never Below
Reasoning	magistral-small	Tier 3
Code	devstral-small	Tier 2
Vision	pixtral-large	Tier 4
General	mistral-small	Tier 2

If budget can't afford the floor, BudgetExceededError is raised instead of silently degrading.

Comparison

Feature	Tramontane	LangGraph	CrewAI	Mistral Agents API
Mistral-native fleet routing	Yes	No	No	Partial
EUR budget ceilings + quality floors	Yes	No	No	No
GDPR PII + Article 30 built-in	Yes	No	No	No
Agentic + deterministic modes	Yes	Yes	Partial	No
Checkpoint/resume	Yes	Yes	No	No
SSE streaming API	Yes	Complex	No	Partial
Voice input (Voxtral)	Yes	No	No	No
MCP client (stdio + SSE)	Yes	No	No	Native
3-agent code review cost	EUR 0.002	~EUR 0.15*	~EUR 0.15*	N/A

*Estimated with GPT-4o equivalent pricing.

The Mistral Fleet

Model	Tier	Best For	EUR/1M in	EUR/1M out
ministral-3b	0	Classification, PII, triage	0.04	0.04
ministral-7b	1	Bulk, extraction, tool calls	0.10	0.10
mistral-small	2	General, multilingual	0.10	0.30
devstral-small	2	All code tasks, SWE	0.10	0.30
magistral-small	3	Reasoning, CoT, planning	0.50	1.50
magistral-medium	3	Deep reasoning	2.00	5.00
devstral-2	4	Complex SWE, monorepo	0.50	1.50
pixtral-large	4	Vision, multimodal, OCR	2.00	6.00
mistral-large	4	Frontier, synthesis	2.00	6.00
voxtral-mini	1	Voice transcription	0.04	0.04

Built-in Pipelines

Pipeline	Agents	Typical Cost	GDPR
code_review	Reviewer, Security Auditor, Writer	EUR 0.002	none
lead_gen_fr	Prospector, Qualifier, Copywriter	EUR 0.001	strict
market_research	Researcher, Analyst, Writer	EUR 0.010	standard
document_analysis	Extractor, Legal Analyst, Summarizer	EUR 0.018	strict

GDPR

Three levels: none (passthrough), standard (detect + log), strict (detect + redact + block).

• PII detection -- dual-mode: regex offline, Ministral-3B online for contextual PII
• French PII -- email, phone (+33), IBAN (FR), NIR, passport
• Right to erasure -- memory.erase_user(user_id), Article 17 compliant
• Audit vault -- append-only, never deletes, Article 30 report generation
• HTTP 451 -- GDPRViolationError returns the correct status code

CLI

tramontane --version                 # show version
tramontane models                    # display the Mistral fleet
tramontane init                      # create database + health check
tramontane run <pipeline.yaml> \
  --input "..." --budget 0.05        # run a pipeline
tramontane serve --port 8080         # start FastAPI server
tramontane audit --run <run_id>      # view audit trail

API Server

tramontane serve --port 8080
# Docs at http://localhost:8080/docs
# SSE streaming: POST /pipelines/run with "stream": true

All responses include X-Tramontane-Version and X-EU-Sovereign: true headers.

Links

• PyPI: https://pypi.org/project/tramontane/
• GitHub: https://github.com/Jesiel-dev-creator/TRAMONTANE
• HF Demo: https://huggingface.co/spaces/BleuCommerce-Apps/TRAMONTANE-demo
• HF Pipelines: https://hf.co/datasets/BleuCommerce-Apps/tramontane-pipelines

Contributing

git clone https://github.com/Jesiel-dev-creator/TRAMONTANE.git
cd TRAMONTANE
uv sync
uv run pytest tests/ -q    # 61 tests
uv run ruff check .        # lint
uv run mypy .              # type check (strict)

License

MIT -- Bleucommerce SAS, Orleans, France