JAOT — Just Another Optimization Tool

Software that decides. How much of each product to make, which orders go on which van, who covers which shift, what to buy and when — decisions with more combinations than anyone can weigh by hand, and a real cost to getting wrong. Describe one in plain language or JSON, and JAOT gives you the best answer, what it is worth, and which limit is the one holding you back.

Live demo → [jaot.io](https://jaot.io) — the reference deployment, running the same images you can build from this repo. Browse the marketplace and the docs without an account; solving needs one.

Quickstart · Architecture · Development · Documentation · License

What it is

You say what you are deciding, what constrains you — capacity, budget, hours, stock — and what "best" means for you: cheapest, fastest, most profit. JAOT turns that into a mathematical model, hands it to an industrial solver, and gives you back the decision in the same terms you asked the question in.

You do not need to know what a MIP is to use it. If you do, nothing is hidden: the model, the solve log, the gap and the full post-solve analysis are all there, and an answer that could only be proven within a bound says so instead of posing as exact.

It is a platform, not a library and not a hosted service — you run it yourself with docker compose up. It comes with a web interface, a REST API, and an MCP server so AI agents can use it too. Five solvers ship with it (SCIP, HiGHS, CBC, GLPK and JAOS) and adding another means writing one adapter.

Free, with no paid tier. No billing, no credits, no upsell — the marketplace is people sharing models, not selling them. Fair use is a set of request limits and solve quotas you configure yourself, and the AI assistant runs on a monthly budget you set, or on your own API key.

Nothing caps the size of your models but your hardware. There is no ceiling on model size, expression length, thread count or solve time — every limit 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.

Solve

• Solver-agnostic core — an OptimizationProblem schema that stays

independent of the solver. Ships SCIP (via PySCIPOpt), HiGHS (via highspy) and JAOS (via its own Python library), and CBC and GLPK as separate command-line programs; an optional Hexaly adapter is bring-your-own-license.

• Compare solvers on the same problem — run one model on several solvers

under identical terms (same time limit, same gap tolerance, one machine, one at a time) and read what each of them actually did: result, objective, best bound, gap, time, nodes and iterations. A solver that cannot express the model says so in its own row instead of leaving a blank. Inside a model's workspace the same thing crosses several datasets with several solvers as a matrix, so the answer is not decided by one month's data.

• 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.

• Standard formats in and out — MPS, LP, CIP and JSON, with a preview before

you commit to a solve.

Understand

• 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.

• Infeasibility diagnosis — when a model cannot be satisfied, get the minimal

conflicting set of constraints instead of a bare INFEASIBLE.

Build and share

• 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).

• 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 and 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.

• 102 templates + 32 problem generators — knapsack, vehicle routing,

scheduling, production planning, portfolio, a full MDPDP-TW formulation, and more. Every template's model is gated: each number in its example must reach the model, the objective must be able to tell two answers apart, and 24 of them are pinned to an optimum worked out by hand.

Integrate and operate

• MCP server — 34 curated tools for AI agents 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.

• REST API v2 — every capability of the UI is an endpoint, authenticated with

a Bearer API key.

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

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

Quickstart

Requirements

• Git and Docker + Docker Compose (the recommended path — everything runs

in containers).

• Or, for a local install without Docker: Python 3.12 and a PostgreSQL

instance.

Run it

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

This brings up PostgreSQL, RabbitMQ, Redis, Qdrant, the API (port 8001), two Celery workers (the general one and the solver-comparison one), Celery beat, and the frontend (port 3000). Both ports bind to 127.0.0.1 only, so nothing is exposed to your network until you put a reverse proxy in front.

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

docker compose exec api python scripts/doctor.py

Open http://localhost:3000 and log in with your SEED_ADMIN_* credentials. See Configuration for the full guide.

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,...}.

Connect an AI agent (MCP)

The MCP server speaks Streamable HTTP at /mcp. Browsing templates and the marketplace needs no key; anything that reads or writes your data takes a Bearer API key:

claude mcp add --transport http jaot http://localhost:8001/mcp \
  --header "Authorization: Bearer <your-api-key>"

Full setup guide (Claude Code, Claude Desktop, opencode, OpenAI Responses API) → docs/getting-started/QUICKSTART.md.

Architecture

flowchart TB
    BROWSER["Browser<br/>Next.js 16 · React 19 · 5 locales"]
    AGENT["AI agent<br/>MCP client"]
    SCRIPT["Script or service<br/>Bearer API key"]

    API["<b>FastAPI</b> · Python 3.12 · :8001<br/>REST /api/v2 · MCP /mcp · WebSocket /ws<br/>auth · solve · studio · marketplace · LLM/RAG · triggers"]

    BROWSER -->|"REST + SSE + WebSocket"| API
    AGENT -->|"MCP, Streamable HTTP"| API
    SCRIPT -->|"REST"| API

    subgraph STORES["State"]
        direction LR
        PG[("PostgreSQL 18<br/>every tenant, one schema")]
        REDIS[("Redis<br/>cache · rate limits")]
        QDRANT[("Qdrant<br/>RAG · 290 docs · 384-dim")]
    end

    API --> PG
    API --> REDIS
    API --> QDRANT
    API -.->|"opt-in, budgeted"| CLAUDE["Anthropic Claude API"]

    API -->|"enqueue a solve"| MQ["RabbitMQ<br/>one queue per solver"]
    MQ --> WORKERS["Celery workers"]
    WORKERS --> ADAPTERS["<b>SolverAdapter protocol</b><br/>app/domains/solver/adapters<br/>SCIP · HiGHS · CBC · GLPK · JAOS<br/>Hexaly, profile-gated"]
    WORKERS -.->|"writes the run"| PG

A modular monolith: one process, one database, boundaries enforced in code. Two bounded contexts are extracted so far — the solver (app/domains/solver/, behind a SolverAdapter protocol) and the JModel compiler (app/domains/dsl/, which may import app.schemas and nothing else). Seven import-linter contracts fail the build if an import crosses a boundary. Adding a solver means writing one adapter — see docs/ARCHITECTURE/OVERVIEW.md.

Development

pip install -r requirements.txt -r requirements-dev.txt
pip install "ruff==0.16.3" "import-linter>=2.1"   # the versions CI pins

pytest                            # backend tests — real PostgreSQL, no mocked DB
ruff check app/ infra/ scripts/ deploy/ tests/    # backend lint (100-char lines)
ruff format --check app/ infra/ scripts/ deploy/ tests/   # CI runs this too
lint-imports                      # the 7 domain boundary contracts

cd frontend                       # Node 24
npm ci
npm run lint                      # frontend lint
npm run test                      # unit tests (vitest)
npm run build                     # the real frontend gate — see the note below
npm run test:e2e                  # end-to-end (playwright, against a prod-like build)

npm run build catches errors that tsc and eslint do not, so run it before you open a PR. Its prebuild step overwrites frontend/src/lib/generated/api.ts from whatever API container is running on port 8001. Start the current API first, or the generated types go stale under you.

ruff and lint-imports run twice: as pre-commit hooks (.pre-commit-config.yaml) and as CI steps, because a hook can be skipped with --no-verify. Both places pin the same versions.

Database migrations:

alembic -c infra/alembic.ini upgrade head
alembic -c infra/alembic.ini revision --autogenerate -m "description"

Conventions, dev setup and the PR checklist are in CONTRIBUTING.md.

Documentation

Built with

JAOT stands on the [SCIP Optimization Suite](https://www.scipopt.org/) (Zuse Institute Berlin) and [HiGHS](https://highs.dev/) — 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, the frontend's API types are regenerated from the backend schema on every push and must match what is committed, 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. It also ships CBC (EPL-2.0) and GLPK (GPL-3.0-or-later) as separate command-line programs, and JAOS (Apache-2.0) as a Python library. 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.