JAOT MCP Server

JAOT — Just Another Optimization Tool

A self-hostable optimization platform. Describe a problem in natural language or JSON, get the optimal solution back — no solver expertise required. Build models with an AI assistant, share them in a marketplace, expose them to AI agents over MCP, or just hit the REST API.

---

What it is

JAOT wraps industrial MIP/LP solvers (SCIP, HiGHS) behind a multi-tenant API, a visual builder, and an LLM formulation assistant. It is a platform, not a library and not a hosted SaaS — you run it yourself with docker compose up.

• Solver-agnostic core — an OptimizationProblem schema that stays

independent of the solver. Ships SCIP (via PySCIPOpt) and HiGHS (via highspy); an optional Hexaly adapter is bring-your-own-license.

• LLM formulation assistant — turn a natural-language description into a

runnable model, grounded in a RAG index over the template library (Qdrant + local sentence-transformers; no data leaves your box except the Claude calls you opt into).

• Analysis that leads with facts — don't just solve, understand. Solutions

come back as decisions (grouped by the model's real index structure, not a wall of x_3_7 = 1 rows) with an honest solve summary (root node / N nodes / time limit + gap) and an exact, solution-based analysis: binding constraints, slack and utilization computed from your actual solution — exact for the integer optimum on every solver. LP sensitivity (shadow prices, reduced costs) stays available with its caveats, and a one-click AI explanation translates the result into plain language grounded strictly in your actual numbers.

• What-if answers measured, not estimated — ask what one more unit would

actually buy you and JAOT perturbs the solved model and solves it again: RHS ranging on the binding constraints (as a tornado chart) and decision regret (what overruling a binary decision costs). Every figure is measured on the real MIP rather than read off an LP relaxation, and a scenario that hits its time limit is reported as a bound, never as an exact number.

• The interface adapts to your solver — each adapter declares what it

supports, so the UI tells you up front what your chosen solver will not give you (a metaheuristic computes no shadow prices) instead of offering a panel that then comes back empty.

• Model studio — one versioned workspace per model: build it on a visual

canvas, with the AI assistant, in a JSON editor, or in the JModel DSL (sets/params — with a mathematical-notation view, draft derivation from flat models, and compile-verified AI generation from a description or a screenshot); analyze health and stats; solve with live progress; commit versions git-style ("what changed + why"), diff and restore them; run the same model against many datasets/scenarios.

• Model marketplace — a free, collaborative gallery: publish a committed

version of your model, and bring any community model into your own studio with one click ("Use in studio" creates your editable, versioned copy). No prices or commissions — authors share; adoption is the metric.

• MCP server — 30 curated tools for AI agents (Claude, etc.) over the Model

Context Protocol: an agent can author a versioned model, solve it, and ask what is saturated, why a model is infeasible, or what one more unit is worth.

• 102 templates + 31 problem generators — knapsack, vehicle routing,

scheduling, production planning, portfolio, a full MDPDP-TW formulation, and more.

• **Multi-tenant auth, admin panel, i18n (en/es/ca/fr/de),

and a Prometheus/Grafana/Alertmanager monitoring stack** — included.

JAOT is free and collaborative — there is no billing, no credits, no paid tier (ADR-008). Fair use is enforced with one set of request limits for the instance and configurable solve quotas; the AI assistant is bounded by a monthly EUR budget with bring-your-own-key support.

Nothing caps the size of your models but your hardware. There is no ceiling on model size, expression length, thread count or solve time — the limits that survived from the hosted-product era are gone, and every remaining one is a setting you own, where 0 means unlimited. A model with two million variables solves if your machine can hold it. A public instance with open registration can set real numbers; a private one need not.

---

Quickstart

git clone https://github.com/avallavall/jaot.git && cd jaot
cp .env.example .env   # includes first-run admin credentials — change the password
docker compose up -d   # migrates, seeds the catalog, creates your admin on first boot

Then check what still needs configuring (SMTP, AI key…):

docker compose exec api python scripts/doctor.py

See Configuration for the full guide.

Open http://localhost:3000 and log in with your SEED_ADMIN_* credentials — or mint an API key and solve over HTTP:

docker compose exec api python scripts/ensure_admin_api_key.py   # prints your API key

curl -X POST http://localhost:8001/api/v2/solve \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  -d '{"name":"test","variables":[{"name":"x","type":"continuous","lower_bound":0,"upper_bound":10}],"objective":{"sense":"maximize","expression":"3*x"},"constraints":[{"name":"c1","expression":"x <= 5"}]}'

Returns {"status":"optimal","objective_value":15.0,...}. Full setup guide → docs/getting-started/QUICKSTART.md.

---

Architecture

┌──────────────────────────────────────────────┐
│  Next.js 16 frontend  (5 locales)             │
└───────────────┬──────────────────────────────┘
                │ REST + SSE + WebSocket
┌───────────────▼──────────────────────────────┐
│  FastAPI (Python 3.12)                        │
│  auth · solve · studio · LLM/RAG ·            │
│  marketplace · triggers · MCP server          │
└──┬─────────┬──────────┬──────────┬────────────┘
   │         │          │          │
┌──▼──┐ ┌────▼────┐ ┌──▼──┐ ┌─────▼─────┐ ┌────────────┐
│ Pg  │ │RabbitMQ │ │Redis│ │  Qdrant   │ │ Anthropic  │
│ 18  │ │+ Celery │ │     │ │ (RAG)     │ │ Claude API │
└─────┘ │ workers │ └─────┘ └───────────┘ └────────────┘
        │ SCIP /  │
        │ HiGHS / │
        │ Hexaly  │
        └─────────┘

A modular monolith: the solver is the first extracted bounded context (app/domains/solver/), behind a SolverAdapter protocol enforced by import-linter contracts. Adding a solver means writing one adapter — see docs/ARCHITECTURE/OVERVIEW.md.

---

Documentation

| Doc | Description | |---|---| | Quickstart | From zero to first solve | | Roadmap | Where the project is heading — now / next / later | | Configuration | Self-hosting config: .env vs admin panel, + the config doctor | | Architecture | System design, components, data model | | Contributing | Dev setup and conventions | | Testing & Quality | Test strategy, coverage, mutation scores | | Disaster Recovery | Incident response runbook | | MDPDP Spec | A worked mathematical formulation |

---

Built with

JAOT stands on the SCIP Optimization Suite (Zuse Institute Berlin) and HiGHS — full attributions in THIRD_PARTY_LICENSES.

Built solo and AI-accelerated. What you can verify rather than take on faith: tests run against real PostgreSQL (no mocked DB), domain boundaries are enforced by import-linter contracts, and every change is gated by lint, tests, and security scans (bandit, pip-audit, npm audit). Details, coverage, and mutation-test scores in Testing & Quality.

Maintained best-effort — monthly issue triage, quarterly dependency/CVE pass. Issues and focused PRs welcome; see CONTRIBUTING.md and SECURITY.md.

---

Citing the solvers

JAOT is powered by SCIP 10 (via PySCIPOpt) and HiGHS. As the SCIP team requests, any work that uses SCIP should acknowledge and cite it. If JAOT helps your research or product, please cite the underlying solvers:

@misc{scip10,
  title        = {The {SCIP} Optimization Suite 10.0},
  author       = {Christopher Hojny and Mathieu Besançon and Ksenia Bestuzheva and Sander Borst and João Dionísio and Johannes Ehls and Leon Eifler and Mohammed Ghannam and Ambros Gleixner and Adrian Göß and Alexander Hoen and Jacob von Holly-Ponientzietz and Rolf van der Hulst and Dominik Kamp and Thorsten Koch and Kevin Kofler and Jurgen Lentz and Marco Lübbecke and Stephen J. Maher and Paul Matti Meinhold and Gioni Mexi and Til Mohr and Erik Mühmer and Krunal Kishor Patel and Marc E. Pfetsch and Sebastian Pokutta and Chantal Reinartz Groba and Felipe Serrano and Yuji Shinano and Mark Turner and Stefan Vigerske and Matthias Walter and Dieter Weninger and Liding Xu},
  year         = {2025},
  howpublished = {Optimization Online preprint, arXiv:2511.18580},
  url          = {https://arxiv.org/abs/2511.18580}
}

@article{achterberg2009scip,
  title   = {{SCIP}: solving constraint integer programs},
  author  = {Achterberg, Tobias},
  journal = {Mathematical Programming Computation},
  volume  = {1},
  number  = {1},
  pages   = {1--41},
  year    = {2009},
  doi     = {10.1007/s12532-008-0001-1}
}

@article{huangfu2018highs,
  title   = {Parallelizing the dual revised simplex method},
  author  = {Huangfu, Qi and Hall, J. A. Julian},
  journal = {Mathematical Programming Computation},
  volume  = {10},
  number  = {1},
  pages   = {119--142},
  year    = {2018},
  doi     = {10.1007/s12532-017-0130-5}
}

---

License

Apache License 2.0 — see also NOTICE. Third-party license attributions are in THIRD_PARTY_LICENSES.