Claude, ChatGPT and other AI clients can now reach outside the chat through the Model Context Protocol (MCP). For insurance pricing, that means an assistant can learn how a model is built, write one, check it and run its tests without leaving the conversation.
This is our honest take on how that goes: what an assistant does well, where it struggles, and the setup that keeps your team in control.
Getting connected
The Swallow MCP server lives at https://api.llow.io/ai/mcp. In Claude on the web or desktop you add it as a custom connector. In Claude Code it's one command, or you can install the Swallow plugin, which adds skills and agents on top. Cursor and other clients take it as a remote server. ChatGPT connects too.
Whichever client you use, it opens a Swallow sign-in page, you paste your client ID and secret keys once, and from then on the client works with a token, never your keys. Step-by-step guides are on our MCP page and in our posts on connecting Claude and ChatGPT.
What the assistant can do
The tools fall into three groups.
Building and testing: docs_swallow_project and schema_swallow_project load the modelling rules and the JSON Schema. starter_swallow_project hands over a small working model for a common line. validate_swallow_project checks a model and points to each problem by its JSON path. test_swallow_project prices every test through the engine. generate_swallow_tests and validate_swallow_tests help build and review a suite. summarise_swallow_project helps with very large models.
Your account: list and open your projects and quotes, read a quote's debug trace, pull quote analytics, show a project's quote form in the chat and run a quote against the latest published version.
Help: dashboard help, a sign-up link for someone new and a health check. These work before you sign in.
Everything is read-only apart from running a quote. No tool creates, edits or publishes a project.
Where it shines
Adapting a starter
Asking an assistant to write a model from a blank page is a gamble. Asking it to take a close example and reshape it to your rating guide works far better. The starter brings a correct structure, and the assistant swaps in your inputs, tables and formulas.
Checking its own work
The assistant can validate after every change and fix structural slips on the spot: a missing default, a reference to a field that doesn't exist, a step in the wrong order. Most errors die there, before a single quote is priced.
Running the suite
With your worked examples saved as tests, the assistant prices them all and tells you which match. When one doesn't, it can read the debug trace, walk back through the steps and name the first one where the numbers go wrong.
Explaining a model
Ask how a premium is built, or why a quote referred, and you get a plain answer drawn from the model and the trace. For people new to a product, this is often the most useful thing it does.
Converting a spreadsheet
The Swallow plugin for Claude Code includes a skill for tracing an Excel rater's formulas and producing an equivalent Swallow model. More on that in replacing a spreadsheet rater.
Where it trips up
Gaps in the rating guide
If the guide doesn't say whether a discount applies before or after the minimum premium, the assistant will pick one. It won't always pick right. Ask it to list every assumption it made, and check each one with the pricing team. Worked examples close most of these gaps.
Big models
AI clients can only take in so much at once, and a model with large rate tables can crowd out the conversation. The assistant can read a project as a step list or a short summary instead of the whole file, and summarise_swallow_project shows it how to shrink a model to its structure. Large tables are often easier to change as a CSV in the dashboard.
Moving the goalposts
The worst thing an assistant can do is change a test's expected result so a failing test passes. Expected results should only change when a person decides they should, for a reason. Say so in your instructions, and review any change to an expected value as closely as a rate change.
Judgement calls
How to treat an unusual risk, whether to refer or decline, what to do about missing data: these are pricing decisions. The assistant can lay out the options. A person should choose.
Keeping people in charge
The setup we recommend follows from how the server is built:
- The assistant drafts and tests. You put it live. When the model is ready, the assistant hands you the JSON. You upload it to a draft in the dashboard, then approve and publish it there.
- Tests run on every change, and failures are reported, not hidden.
- Approval follows your team's roles. Every version moves from draft to approved to published, and each step is logged with who did it and when.
- Read the report, then the change. Look at the test results and summary first, then check the specific edits.
Used like this, the assistant takes on the slow, repetitive parts of model building and the pricing decisions stay with your team. The model format is what makes the review quick, which we explain in why pricing models should be human readable.
Try it
Connect your client to https://api.llow.io/ai/mcp and ask it to build from a starter. If you don't have an account yet, talk to our team and we'll get you set up.
