# "Quickstart: claude.ai"

Add the DSAIL connector to claude.ai, describe one policy, and get a checked result — three steps, no code.

For a person working in claude.ai on the web, desktop or iOS. Turn a written policy into rules a program can check without leaving the conversation.

## 1. Add the connector

Settings → Connectors → **Add custom connector**. Name it `DSAIL`, and for the
URL enter:

```text
https://agents.jaxon.ai/mcp
```

A consent page from Jaxon opens with one **Approve** button. There is no account
to create. After approving, the connector shows as connected.

## 2. Describe the policy

In a new conversation, with the connector enabled, paste the policy in English:

> Turn this into rules a program can check: expenses over 75 USD need an
> itemised receipt, and no single claim may exceed 5,000 USD.

Claude restates the policy and asks you to confirm it before writing any DSAIL.
Confirm. It compiles the ruleset and opens the **review widget** inline: the
source, the questions a document will be asked (the claim manifest), and an
exercise panel.

## 3. Get a result

In the widget's exercise panel enter values — `120 USD` for the amount, `no` for
the receipt — and run the check. Each assertion answers in its own row:

```text
receipt_over_75   FALSE   counterexample: amount = 120 USD, has_receipt = false
within_cap        TRUE
```

That is the whole loop. When you are satisfied, **Approve** in the widget: the
approval is recorded against the exact hash of the source you looked at, and
the ruleset is saved under the name you give it. Later conversations on this
connector list it under `dsail_list_rulesets`.

## What to expect

- Claude will not invent a result. Every assertion is reported as `TRUE`,
  `FALSE`, `UNKNOWN` or `AMBIGUOUS`, attributed to the rules; there is no overall
  verdict and it is told not to produce one.
- If a value cannot be determined, say so: unknown is an answer, not a guess,
  and the assertion that needed it answers `UNKNOWN` rather than guessing.
- Rulesets saved on this connector are visible to every conversation that uses
  it, and to the same connector URL added by anyone else until per-user identity
  ships. Do not put a customer's policy on the shared evaluation deployment.

## If it goes wrong

Every error the connector returns carries a `docs` URL to the page that resolves
it; [Errors](../reference/errors.md) lists them all. If the widget does not
render, the tool result text carries the same payload as JSON.

Next: [Quickstart: Claude Code](claude-code.md) for the same ruleset in a
repository, or the [MCP tools reference](../reference/mcp-tools.md).