> ## Documentation Index
> Fetch the complete documentation index at: https://vaquill.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Claude Code Plugin

> The primary-law plugin for Claude Code: ten research commands, a US statute research skill, and the Vaquill MCP server in one install.

The **primary-law** plugin packages the Vaquill MCP server together with ten research commands and a skill that teaches Claude how to use them. It installs into Claude Code from the Vaquill plugin marketplace.

Source: [github.com/Vaquill-AI/vaquill-plugin](https://github.com/Vaquill-AI/vaquill-plugin).

<Note>
  **The plugin is Claude Code only.** It does not run in Cowork, claude.ai, or Claude Desktop. For those, use the [Claude Connector](/docs/integrations/mcp/claude-connector), which needs no environment variable and no local shell.
</Note>

## Install

<Steps>
  <Step title="Add the marketplace">
    ```bash theme={"theme":"github-dark"}
    /plugin marketplace add Vaquill-AI/vaquill-plugin
    ```
  </Step>

  <Step title="Install the plugin">
    ```bash theme={"theme":"github-dark"}
    /plugin install primary-law
    ```
  </Step>

  <Step title="Set your API key where a new shell will see it">
    The plugin reads `VAQUILL_API_KEY` from the **process** environment. Create a key at [app.vaquill.ai/settings](https://app.vaquill.ai/settings) under **API Keys**, then:

    ```bash theme={"theme":"github-dark"}
    echo 'export VAQUILL_API_KEY="vq_key_YOUR_KEY"' >> ~/.zshenv
    chmod 600 ~/.zshenv
    ```

    **`~/.zshrc` is not enough.** It is sourced only for interactive shells, and Claude Code's child processes are not always interactive. `~/.zshenv` is sourced for every zsh invocation. On bash, use `~/.bash_profile` plus `~/.bashrc`. A bare `export` typed into a terminal dies with that shell.

    Check it reaches a **new** shell, which is what actually matters:

    ```bash theme={"theme":"github-dark"}
    zsh -c 'echo ${VAQUILL_API_KEY:+PRESENT}'    # must print PRESENT
    ```

    Use the `:+` form. `${VAR:-...}` prints the key itself into your transcript.
  </Step>

  <Step title="Fully quit and relaunch Claude Code">
    Environment is inherited **once, at process start**. None of these pick up a new variable: `/reload`, MCP servers reconnecting, a new session in the same window, or reinstalling the plugin.

    Quit Claude Code itself (`/exit` or Ctrl+D) and run `claude` again from a shell that has the variable. `pgrep -f claude` must return a **different** pid than before.
  </Step>

  <Step title="Make one live call">
    See [Verify it works](#verify-it-works).
  </Step>
</Steps>

Never write the key into `.mcp.json`, a committed `.env`, or any file under version control. The variable indirection exists precisely so the key never lands in the repo.

## Verify it works

<Warning>
  **A green connection proves nothing.** The server completes the handshake and lists all 25 tools with no `Authorization` header at all, with an unexpanded `${VAQUILL_API_KEY}` template, and with an invalid key. `claude mcp list` showing `Connected`, and a full tool list in `/mcp`, are both consistent with a completely broken credential.
</Warning>

The only real smoke test is one live call, and the cheapest is free:

> "Use Vaquill to list statutes coverage for California."

`list_statutes_coverage` costs **0 credits**. A 401 means the key never reached the server. A normal response means setup is done.

## Commands

Ten slash commands, each written to state its own caveats rather than leave you to infer them.

| Command             | What it does                                                                                                |
| ------------------- | ----------------------------------------------------------------------------------------------------------- |
| `/research`         | Find the controlling US law on a question, answering only from the text the tools return                    |
| `/verify-citations` | Verify every citation in a file, selection, or pasted text against the official source text                 |
| `/as-of`            | Read a section as it stood on a given date, with the observation caveat stated honestly                     |
| `/changed`          | Sweep a corpus or jurisdiction for what has changed since a date                                            |
| `/authority`        | Trace the chain between a statute and the regulations implementing it, in either direction                  |
| `/compare-states`   | Compare how different states word the same rule                                                             |
| `/coverage`         | Report what Vaquill actually holds for a jurisdiction, before you read an empty result as an absence of law |
| `/guidance`         | Find federal agency guidance on a topic, and say what weight it does and does not carry                     |
| `/adjudications`    | Search federal administrative adjudications, with the overruling caveat stated up front                     |
| `/rulemaking`       | Track a Federal Register rulemaking: what was proposed, what became final, when comments close              |

## The research skill

The plugin also ships a `us-statute-research` skill. Claude loads it when a question calls for primary law, and it carries the judgement that the tool descriptions cannot: how to scope a search, when a citation needs resolving before it can be read, and why an empty result is more often a malformed `act_id` than a gap in coverage.

## Key caps and rotation

* **Active key caps are low.** Business allows 5 keys; every other plan allows 2. The console's own Playground mints a key when it cannot reuse one, so one integration key plus Playground use already reaches the cap of 2.
* **Rotation differs by plan.** Business can create the new key before revoking the old one, so there is no gap. Other plans must revoke first and accept downtime.

A key provisioned by the [Claude Connector](/docs/integrations/mcp/claude-connector) does **not** count against this cap, and does not appear in your dashboard key list.

## Troubleshooting

| Symptom                        | Cause                                                                                       |
| ------------------------------ | ------------------------------------------------------------------------------------------- |
| 401 on every tool              | Key missing, malformed, or the `Bearer ` prefix was dropped. The key alone is rejected      |
| 401 right after revoking a key | Key validation is cached about 60 seconds, so a revoked key can work briefly                |
| 401 with a healthy `/mcp`      | The variable never reached the process. Check `pgrep -f claude` changed after your relaunch |
| 402                            | Out of credits. Top up in the console                                                       |
| 429                            | Rate limited. Honour `Retry-After`. `X-RateLimit-Limit` under-reports on Business keys      |
| 503                            | Backend unavailable. The call was refunded. Retry later                                     |
| Tools missing entirely         | The server failed to connect. Check `/mcp` before assuming the key is wrong                 |
| Nothing works in Cowork        | Expected. Use the [connector](/docs/integrations/mcp/claude-connector)                           |

An empty result set is not an authentication problem. Since an outage returns 503 with a refund, an empty 200 is a real answer.

## Attribution

**On the Pro plan, attribution is required**: a visible "Legal data by Vaquill AI" credit, linked to the site, wherever the data is surfaced. Business waives it. If you are building something customer-facing on Pro, this is a licence term rather than a suggestion.