SwarmLabs Engine Kit

Open-source client SDK, MCP server, and Skill catalog for the SwarmLabs physics-informed multi-agent scientific experiment automation engine.

SwarmLabs is an open infrastructure that turns multidisciplinary scientific experiments (chemistry, energy, materials, environment, biology, pharma, quantum computing, brain science, CFD, structural mechanics, …) into callable, auditable physics-informed predictions served over a small HTTP API. This repository is the open-source developer toolkit around that engine — it does not contain the proprietary physics models, only the interface, reference integrations, and honesty-first contracts that let any agent or application consume the engine safely.

Why this exists

Scientific-AI tooling is flooded with demos that return confident-but-fake numbers. SwarmLabs takes the opposite stance — honesty-first:

• Every prediction reports an uncertainty and, where applicable, an epistemic_uncertainty (how far the request is from validated parameter space).

• Validation is split into empirical_validated (matched to real published experiment data) and literature_validated (matched to a real paper), never a single inflated "validated" count.

• The v2/pi (physics-informed) endpoint returns the real physics model — it does not return a constant placeholder.

This kit makes those guarantees easy to build on top of, from a Python script, an MCP-compatible agent runtime, or a no-code agent builder.

Related MCP server: yellow-pages

What's inside

Install

pip install swarmlabs-engine-kit

Or use it directly from source:

git clone https://github.com/lm203688/swarmlabs-engine-kit.git
cd swarmlabs-engine-kit
pip install -e .

Quick start

from swarmlabs_engine import SwarmLabsClient

# Point the client at YOUR deployed SwarmLabs engine endpoint.
client = SwarmLabsClient(base_url="https://your-swarmlabs-engine.example.com")

# List available engines
info = client.list_engines()
print(info["engines"], "engines across", info["physics_models"], "physics models")

# Run a real physics-informed prediction (e.g. H2 dissociation via VQE)
result = client.run("vqe_h2", {"bond_length_A": 0.74})
print(result["result"], "+/-", result["uncertainty"])

See examples/quickstart.py for a fuller walkthrough.

Engine API surface (summary)

All endpoints are served under /api/v2/.

The engine deployment URL is provided by you. This kit is the client side and works against any SwarmLabs engine deployment that exposes the /api/v2/ surface.

MCP integration

The reference MCP server turns engine calls into standard MCP tools so any MCP-aware agent (Claude Desktop, Cursor, custom runtimes) can use SwarmLabs as a trusted scientific-compute tool:

python -m mcp.server --base-url https://your-swarmlabs-engine.example.com

It exposes swarmlabs_run, swarmlabs_list, and swarmlabs_sweep tools with explicit input schemas and the same honesty-first result contract.

Skill catalog

skills/skill_catalog.json describes the 4 core Skills that compose a SwarmLabs multi-agent workflow:

1. PhysicsPredictSkill — call the real engine.

2. ExperimentDesignSkill — turn a goal into a parameter plan.

3. ActiveLearningSkill — pick the next most informative experiment point.

4. ReportGenerationSkill — emit an auditable report.

These are the building blocks of the SwarmLabs "Planner → Executor → Verifier" agent loop.

License

MIT.