marimo-flow 0.3.1

Reactive marimo notebooks for ML with MLflow tracking, PINA physics-informed neural networks, and a multi-agent team on pydantic-graph + Ollama Cloud.

Marimo Flow 🌊

---

Like marimo algae drifting in crystal waters, your data flows and evolves – each cell a living sphere of computation, gently touching others, creating ripples of reactive change. In this digital ocean, data streams like currents, models grow like organic formations, and insights emerge naturally from the depths. Let your ML experiments flow freely, tracked and nurtured, as nature intended.

https://github.com/user-attachments/assets/3bc24463-ff42-44a7-ae61-5d500d29688c

Why Marimo Flow is Powerful 🚀

Marimo Flow combines reactive notebook development with AI-powered assistance and robust ML experiment tracking:

• 🤖 AI-First Development with MCP: Model Context Protocol (MCP) integration brings live documentation, code examples, and AI assistance directly into your notebooks - access up-to-date library docs for Marimo, Polars, Plotly, and more without leaving your workflow
• 🔄 Reactive Execution: Marimo's dataflow graph ensures your notebooks are always consistent - change a parameter and watch your entire pipeline update automatically
• 📊 Seamless ML Pipeline: MLflow integration tracks every experiment, model, and metric without breaking your flow
• 🎯 Interactive Development: Real-time parameter tuning with instant feedback and beautiful visualizations

This combination eliminates the reproducibility issues of traditional notebooks while providing AI-enhanced, enterprise-grade experiment tracking.

Features ✨

🤖 AI-Powered Development (MCP)

• Model Context Protocol Integration: Live documentation and AI assistance in your notebooks
• Context7 Server: Access up-to-date docs for any Python library without leaving marimo
• Marimo MCP Server: Specialized assistance for marimo patterns and best practices
• Local LLM Support: Ollama integration for privacy-focused AI code completion

📊 ML Development Workflow

• 📓 Reactive Notebooks: Git-friendly .py notebooks with automatic dependency tracking
• 🔬 MLflow Tracking: Complete ML lifecycle management with model registry
• 🎯 Interactive Development: Real-time parameter tuning with instant visual feedback
• 💾 SQLite Backend: Lightweight, file-based storage for experiments

🧠 PINA — composition-first (no hardcoded PDE factories)

• compose_problem(ProblemSpec): agents emit typed EquationSpec + SubdomainSpec + ConditionSpec and the composer compiles a pina.Problem subclass at runtime (sympy.lambdify + pina.operator.grad/laplacian). No ProblemKind enum — any PDE sympy can express is reachable.
• Inverse problems: UnknownParameterSpec + data-fitting ObservationSpec → composer wires a pina.LearnableParameter and routes a 3-arg residual.
• Mesh geometry: MeshSpec + meshio + MeshDomain barycentric sampling (tri/tetra/quad/hex). SubdomainSpec.mesh_ref picks cell groups.
• 3D visualisation: core/viz3d.py via plotly (Mesh3d, Volume, Scatter3d, Isosurface) — no 150 MB VTK stack.
• Design optimisation: OptimizationPlan + DesignVariableSpec + ConstraintSpec driving Optuna TPE or scipy SLSQP with penalty / augmented-Lagrangian handling.
• Stochastic + non-local: NoiseSpec (white / colored / fbm) wraps the residual; fractional Laplacian via Riesz-kernel Monte-Carlo quadrature.
• MPC: marimo_flow.control — rolling-horizon scipy SLSQP on a trained PINN surrogate.
• Walrus foundation model adapter for Poisson-class problems.

🐳 Deployment

• Docker: docker-compose setup with CPU, CUDA, and XPU image variants
• 📚 MCP-Powered Docs: Live documentation via Context7 and Marimo MCP servers

Quick Start 🏃‍♂️

With Docker (Recommended)

# Clone repository
git clone https://github.com/synapticore-io/marimo-flow.git
cd marimo-flow

# Build and start services
docker compose -f docker/docker-compose.yaml up --build -d

# Access services
# Marimo: http://localhost:2718
# MLflow: http://localhost:5000

# View logs
docker compose -f docker/docker-compose.yaml logs -f

# Stop services
docker compose -f docker/docker-compose.yaml down

Docker Image Variants

Variant	Image Tag	Use Case
CPU	ghcr.io/synapticore-io/marimo-flow:latest	No GPU (lightweight)
CUDA	ghcr.io/synapticore-io/marimo-flow:cuda	NVIDIA GPUs
XPU	ghcr.io/synapticore-io/marimo-flow:xpu	Intel Arc/Data Center GPUs

# NVIDIA GPU (requires nvidia-docker)
docker compose -f docker/docker-compose.cuda.yaml up -d

# Intel GPU (requires Intel GPU drivers)
docker compose -f docker/docker-compose.xpu.yaml up -d

Local Development

# Install dependencies
uv sync

# Start MLflow server (in background or separate terminal)
uv run mlflow server \
  --host 0.0.0.0 \
  --port 5000 \
  --backend-store-uri sqlite:///data/mlflow/db/mlflow.db \
  --default-artifact-root ./data/mlflow/artifacts \
  --serve-artifacts

# Start Marimo (in another terminal)
uv run marimo edit examples/

Example Notebooks 📚

All notebooks live in examples/ and can be opened with uv run marimo edit examples/<file>.py.

• 01_pina_poisson_solver.py – Solve the Poisson equation with baseline PINNs or the Walrus foundation model. Training is tracked in MLflow with integrated Optuna sweep analytics and experiment history. Uses marimo_flow.core directly.
• 02_provenance_dashboard.py – Review surface over the DuckDB provenance store. Five tables (tasks, experiments, agent decisions, validation verdicts, handoffs) side-by-side plus 3D preset preview. Uses marimo_flow.agents.services.ProvenanceStore.
• 03_navier_stokes_3d_cavity.py – 3D lid-driven cavity composed end-to-end from a declarative ProblemSpec (no hardcoded NS factory). Sliders for viscosity, lid speed, collocation-point count, hidden width, epochs. Renders the spatial domain + mid-plane velocity slice with plotly. Uses the composition-first services/composer.py.
• 04_mpc_heat_rod.py – Closed-loop MPC on a 1D heat-rod PINN surrogate. Trains a small surrogate via the composer, then drives a rolling-horizon scipy-SLSQP MPC loop from marimo_flow.control toward a user-set temperature setpoint.
• lab.py – PINA multi-agent team chat demo (see PINA Multi-Agent Team). Requires Ollama running locally.

Project Structure 📁

marimo-flow/
├── examples/                         # Marimo notebooks
│   ├── 01_pina_poisson_solver.py        # Poisson PINN demo (uses core/)
│   ├── 02_provenance_dashboard.py       # DuckDB review + 3D preset preview
│   ├── 03_navier_stokes_3d_cavity.py    # 3D NS lid-driven cavity (composer)
│   ├── 04_mpc_heat_rod.py               # Closed-loop MPC via scipy SLSQP
│   └── lab.py                           # PINA team chat demo (uses agents/)
├── src/marimo_flow/                  # Installable package
│   ├── core/                         # PINA solvers, training, plotly viz3d
│   ├── control/                      # Rolling-horizon MPC (scipy SLSQP)
│   └── agents/                       # Multi-agent team (pydantic-graph + MLflow)
│       ├── nodes/                    # TriageNode, RouteNode, specialists, ValidationNode
│       ├── schemas/                  # TaskSpec / ProblemSpec / ObservationSpec /
│       │                             #   MeshSpec / OptimizationPlan / ControlPlan / …
│       ├── toolsets/                 # FunctionToolset per role (incl. design, control)
│       └── services/                 # composer, mesh_domain, design aggregator,
│                                     #   provenance (DuckDB 13 tables), experiments
├── tests/                            # Pytest suite (216 passing, 1 xfailed)
├── docs/                             # Project documentation (see docs/INDEX.md)
├── docker/                           # Dockerfiles + compose (CPU, CUDA, XPU)
├── data/mlflow/                      # MLflow storage (artifacts, db)
└── pyproject.toml                    # Dependencies

Two Workflows

Workflow	Import	Use Case
Classic (core/)	from marimo_flow.core import ...	You know the PDE, pick a solver, log to MLflow. See examples/01_pina_poisson_solver.py.
Agents (agents/)	from marimo_flow.agents import lead_chat, FlowDeps	Describe the problem in natural language; a multi-agent team composes Problem + Model + Solver. See examples/lab.py.

Both write to the same MLflow backend (data/mlflow/). The two packages do not depend on each other — pick whichever matches the task.

The marimo_flow.core Package

from marimo_flow.core import (
    ProblemManager,           # Define PDE problems and domains
    SolverManager,            # Configure PINN / SAPINN solvers
    FoundationModelAdapter,   # Walrus foundation model adapter
    create_model_for_problem, # Build a PINA neural-network model for a Problem
    train_solver,             # Run training via PINA's Trainer
    build_optuna_history_figure,
    build_optuna_param_importance_figure,
    build_optuna_parallel_figure,
    build_trials_scatter_chart,
    study_trials_dataframe,
)

MCP (Model Context Protocol) Integration 🔌

marimo and AI-assisted IDEs share MCP servers for live documentation and notebook operations. For the full configuration reference see docs/mcp-setup.md.

marimo (in-notebook AI)

Pre-configured in .marimo.toml:

[mcp]
presets = ["marimo", "context7"]

[mcp.mcpServers.mlflow]
command = "mlflow"
args = ["mcp", "run"]

[mcp.mcpServers.mlflow.env]
MLFLOW_TRACKING_URI = "http://localhost:5000"

[ai.ollama]
model = "gpt-oss:20b-cloud"
base_url = "http://localhost:11434/v1"

The Docker container uses a separate docker/.marimo.toml without MCP presets — containerized sessions run only the notebook UI; MCP servers run on the host and are reached over host.docker.internal.

VS Code / Claude Code

Four MCP servers configured in .vscode/mcp.json:

Server	Purpose
marimo	Notebook inspection, linting (HTTP on port 2718)
mlflow	Trace search, feedback, evaluation (stdio via mlflow mcp run)
context7	Live library documentation (stdio via npx)
serena	Semantic code search (stdio via uvx)

Start the marimo MCP server (required for the marimo tool):

uv tool install "marimo[lsp,recommended,sql,mcp]>=0.18.0"
marimo edit --mcp --no-token --port 2718 --headless

Claude Code skills & hooks

Three domain skills in .claude/Skills/ (marimo, mlflow, pina) provide expert guidance and pre-resolved context7 library IDs (/marimo-team/marimo, /mlflow/mlflow, /mathlab/pina).

Automated cross-platform hooks in .claude/settings.json:

Hook	Trigger	Action
PostToolUse	Edit/Write .py files	Auto-format with ruff
PreToolUse	Edit uv.lock	Block (protection)

Configuration ⚙️

Environment Variables

Docker setup (configured in docker/docker-compose.yaml):

• MLFLOW_BACKEND_STORE_URI: sqlite:////app/data/mlflow/db/mlflow.db
• MLFLOW_DEFAULT_ARTIFACT_ROOT: /app/data/mlflow/artifacts
• MLFLOW_HOST: 0.0.0.0 (allows external access)
• MLFLOW_PORT: 5000
• OLLAMA_BASE_URL: http://host.docker.internal:11434 (requires Ollama on host)

Local development:

• MLFLOW_TRACKING_URI: defaults to sqlite:///data/mlflow/db/mlflow.db (lead agent auto-creates data/mlflow/{db,artifacts}/ and pins the marimo-flow experiment's artifact root). Set to http://localhost:5000 if you run a separate MLflow server.

Docker Services

The Docker container runs both services via docker/start.sh:

• Marimo: Port 2718 - Interactive notebook environment
• MLflow: Port 5000 - Experiment tracking UI

GPU Support: Use docker-compose.cuda.yaml for NVIDIA GPUs or docker-compose.xpu.yaml for Intel GPUs. The default docker-compose.yaml is CPU-only.

PINA Multi-Agent Team (marimo_flow.agents) 🧑‍🚀🧑‍🚀🧑‍🚀

Reactive multi-agent team that orchestrates PINA workflows via pydantic-graph, backed by MLflow for tracing + persistence, exposed via marimo's chat UI and optionally as A2A and AG-UI ASGI servers.

from marimo_flow.agents import lead_chat, FlowDeps
import marimo as mo

deps = FlowDeps()  # uses sqlite:///data/mlflow/db/mlflow.db by default
chat = mo.ui.chat(lead_chat(deps=deps))
chat

Roles (each loads its .claude/Skills/<name>/SKILL.md as instructions= where applicable):

• triage — parses free-form user intent into a typed TaskSpec (start node)
• notebook — marimo MCP cell ops (skills: marimo, marimo-pair)
• problem — defines a PINA Problem from an open spec (skill: pina-problem)
• model — designs a neural architecture for the problem (skill: pina-model)
• solver — wires Solver + Trainer config (skill: pina-solver)
• training — runs pina.Trainer.fit() via the training toolset (skill: pina-training)
• validation — grades the run against task_spec.constraints and records a ValidationReport with an accept/retry/escalate/reject verdict
• mlflow — MLflow MCP tracking + registry (skill: mlflow)
• lead — chat/A2A/AG-UI front-end; wraps the whole graph as one tool

TriageNode runs first and produces the TaskSpec. RouteNode then dispatches between specialists, emits a HandoffRecord on every dispatch, and short-circuits to End when the validation verdict is escalate / reject (SPEC §13 HITL).

Typed specs + provenance (SPEC §8, §12): Every graph run builds typed ProblemSpec / ModelSpec / SolverPlan / RunConfig on FlowState and mirrors them — plus AgentDecision, HandoffRecord, ValidationReport, ExperimentRecord, ArtifactRef, and lineage edges — into a DuckDB provenance store (./provenance.duckdb by default, or MARIMO_FLOW_PROVENANCE_DB). MLflow still owns the binary artifacts; DuckDB owns the queryable index. DuckDB 1.5.2 ships transitively via marimo[sql].

from marimo_flow.agents.services import ProvenanceStore

store = ProvenanceStore("provenance.duckdb")
print(store.query("SELECT title, verdict FROM tasks t "
                  "LEFT JOIN validation_reports v USING (task_id) "
                  "ORDER BY t.created_at DESC LIMIT 10"))

See examples/02_provenance_dashboard.py for a marimo review surface.

Models: provider-prefixed specs ("<provider>:<model>") resolved through pydantic-ai's infer_model. Defaults in marimo_flow.agents.deps.DEFAULT_MODELS all point at Ollama Cloud (http://localhost:11434/v1, :cloud-suffixed tags).

Override per role either via config.yaml at the repo root (see config.yaml.example) or with MARIMO_FLOW_MODEL_<ROLE>=<spec> env vars. Any provider in the pydantic-ai catalogue works — openai, anthropic, groq, mistral, google-gla, bedrock, together, fireworks, openrouter, deepseek, cerebras, xai, ollama, huggingface, ...

Standalone servers:

uv run python -m marimo_flow.agents.server.a2a    # A2A on :8000
uv run python -m marimo_flow.agents.server.ag_ui  # AG-UI on :8001

See examples/lab.py for the full demo notebook.

Contributing 🤝

We welcome contributions! Please see our Contributing Guidelines for details on:

• Development setup and workflow
• Code standards and style guide
• Testing requirements
• Pull request process

Quick Start for Contributors:

1. Fork the repository
2. Create a feature branch: git checkout -b feature-name
3. Make your changes following the coding standards
4. Test your changes: uv run pytest
5. Submit a pull request

See CONTRIBUTING.md for comprehensive guidelines.

Changelog 📋

See CHANGELOG.md for a detailed version history and release notes.

Current Version: 0.3.1

License 📄

This project is licensed under the MIT License - see the LICENSE file for details.

---

Built with ❤️ using Marimo and MLflow