Use ReliaSim® From Claude or ChatGPT
The Assistant Asks the Engine, So the Numbers Come From the Simulation

Ask a general AI assistant how much a buffer would add to your line and it will give you a confident number with no model behind it. Connect it to ReliaSim and it can look the answer up instead. It calls ReliaSim’s tools, gets results the simulation engine produced, and quotes them.

Why connect an assistant to a simulation

A language model is good at explaining, summarizing and answering follow-up questions. It can’t simulate your line, because it has no model of the line to run. When it answers a what-if question on its own, the number is a plausible guess. Why an AI can’t answer what-if goes into the reasons.

The Model Context Protocol (MCP) lets an assistant call outside tools during a conversation. ReliaSim publishes an MCP server, so the assistant can ask for a bottleneck analysis or a buffer sweep and get back results from the engine.

What a conversation looks like

Ask about your line: where it’s held back, what a bigger buffer would do, which fix pays first. You drive the conversation and the model does the math. From one conversation you can:

• See how the line is set up. Which machines, where the buffers sit and how often each one stops, in plain language.
• Run simulations on demand. Uptime by machine, line efficiency, and the range across many runs, in seconds.
• Watch the buffers. How full each one runs, how often it empties or fills, and what that costs in output.
• Get diagrams and reports. Line diagrams and reports you can hand to your team.

Here is one exchange about the 90-day run of the bottling line.

ReliaSim AI · Bottling Line

You

The model matches our 54.3% OEE. Which fix should we make first?

Not the one the Loss Tree ranks first. Labeler Misalignment costs 6.79% and Filler Micro Stop costs 6.72%, nearly a tie. Removing each one in the model, the Labeler fix gives back +5.0 points and the Filler fix gives back +8.1, 62% more, because the Filler’s short stops ripple into every machine after it. Start with the Filler.

You

Could a buffer after the Filler do the same job without touching the machine?

It can soak up some of those short stops. Run the buffer trade-off on this line against the Filler fix and the model shows which one gives back more output for the money. Both answers come from the same playing field.

Every response comes from the model’s own engine runs. It works with Claude, Claude Code and any MCP-compatible assistant.

What the ReliaSim server can answer

The public server at https://reliasim.com/mcp covers the eight reference bottling-line models in the Sandbox. Its tools return saved results from ReliaSim engine runs on those models:

• Bottleneck analysis. Which machine limits the line, with downtime by failure mode, throughput and OEE.
• Gain and loss. What the line gets back when each failure mode is removed one at a time, which often differs from what that mode appeared to cost.
• Buffer sweeps. Throughput as a buffer grows, and where the gains flatten.
• Model facts and comparisons. Each line’s layout, rates and failure data, and two models side by side.
• Concepts. Plain explanations of constraints, buffers, interrupts and the method.

Each tool tells the assistant to quote its numbers as returned. The server doesn’t need an API key or an account.

Connect from Claude

In Claude on the web or the desktop app, add a custom connector in your connector settings and paste the server URL:

https://reliasim.com/mcp

In Claude Code, add it from the terminal:

claude mcp add --transport http reliasim https://reliasim.com/mcp

Other MCP clients, such as Cursor, take the same URL. Custom connectors depend on your plan and on your workspace settings.

Connect from ChatGPT

ChatGPT can use a remote MCP server as a custom connector where your plan and workspace allow it. Add a connector, give it the same URL, and choose no authentication. If you don’t see the option, your workspace admin may need to turn on custom connectors first.

Questions to try

• “On the Buffer Options LEDS line, which failure mode is worth fixing first?”
• “On the Buffer Options line, how big should a buffer be before the gain flattens?”
• “Compare the constraint-level and LEDS-level versions of the same line.”
• “What’s the difference between a constraint and a converter?”

If the answer includes a number, ask the assistant which tool it came from. A good answer names the tool and the model.

Stop-data statistics from ReliaStats

A second server covers the statistics side: Weibull shape, availability, distribution choice and how to read validation bands. It lives at https://reliasim.com/reliastats/mcp/v1 and connects the same way. See ReliaStats for the browser tools it pairs with.

Or use the assistant in the Sandbox

Every demo in the Sandbox has an assistant beside it that answers from that demo’s own results, including the supply chain networks. It’s the quickest way to see the approach before you connect anything. What each Sandbox tier covers. For your own line or network, the models are built from your data as a project. How a ReliaSim project runs shows the steps.