Vaquill

vaquill-mcp

MCP server for Vaquill legal research API. Covers US federal and 50-state primary law: USC, CFR, state statutes and regulations, state and US constitutions, court rules, the Federal Register, executive orders, and agency guidance. Search primary law, resolve statutory citations, browse the hierarchy, and ground answers in official sources, all from your AI tools.

[Discord]

Quick Start

Prerequisites

Sign up at vaquill.ai to get your API key.

Claude.ai (Web)

No installation needed. Add as a remote MCP server from Customize > Connectors > Add custom connector:

Option A: Bearer token (recommended)

Open the Request headers section of the same dialog and add the credential there:

URL:            https://mcp.vaquill.ai/mcpHeader name:    AuthorizationHeader value:   Bearer vq_key_your_key_here

Claude sends the value exactly as you type it and adds no scheme of its own, so the value field holds Bearer vq_key_... and not Authorization: Bearer vq_key_....

Option B: API key in the URL (compatibility)

https://mcp.vaquill.ai/s/vq_key_your_key_here

Still live, and still the only form that works where a client cannot send a header at all. Prefer Option A everywhere else: a credential in a URL is recorded in server logs, proxies and browsing history, which is why Anthropic's connector guidance advises against it and the MCP specification prohibits it.

Available on Claude Pro, Max, Team, and Enterprise plans. The Request headers section is in beta and is enabled per account, it accepts a short allowlist of header names, and it holds at most four. On Team and Enterprise an owner adds the connector under Organization settings first.

Claude Desktop

Open the config file from Settings > Developer > Edit Config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

{  "mcpServers": {    "vaquill": {      "command": "uvx",      "args": ["vaquill-mcp"],      "env": {        "VAQUILL_API_KEY": "vq_key_your_key_here"      }    }  }}

For Indian legislation, add --jurisdiction IN. One process serves one corpus, so run two entries if you want both:

{  "mcpServers": {    "vaquill": {      "command": "uvx",      "args": ["vaquill-mcp"],      "env": { "VAQUILL_API_KEY": "vq_key_your_key_here" }    },    "vaquill-india": {      "command": "uvx",      "args": ["vaquill-mcp", "--jurisdiction", "IN"],      "env": { "VAQUILL_API_KEY": "vq_key_your_key_here" }    }  }}

Quit Claude completely and reopen it. It does not reload the file.

Claude Code

Remote (no install):

claude mcp add --transport http --scope user vaquill \  https://mcp.vaquill.ai/mcp \  --header "Authorization: Bearer vq_key_your_key_here"

Local (via uvx):

claude mcp add --scope user vaquill -e VAQUILL_API_KEY=vq_key_your_key_here \  -- uvx vaquill-mcp

--scope user registers the server for every project; the default is the current directory only. -e stores the key with the registration, so it survives Claude Code being launched from an IDE, which an export in your shell profile does not. Both --header and -e are variadic, so they have to come after the server name.

Cursor

Edit ~/.cursor/mcp.json for every project, or .cursor/mcp.json for one.

Remote:

{  "mcpServers": {    "vaquill": {      "url": "https://mcp.vaquill.ai/mcp",      "headers": {        "Authorization": "Bearer vq_key_your_key_here"      }    }  }}

Local (via uvx):

{  "mcpServers": {    "vaquill": {      "command": "uvx",      "args": ["vaquill-mcp"],      "env": {        "VAQUILL_API_KEY": "vq_key_your_key_here"      }    }  }}

VS Code (Copilot)

Add to .vscode/mcp.json. For every project instead of one, run the MCP: Open User Configuration command and use the same shape there.

Remote:

{  "inputs": [    {      "type": "promptString",      "id": "vaquill-authorization",      "description": "Authorization header value",      "password": true    }  ],  "servers": {    "vaquill": {      "type": "http",      "url": "https://mcp.vaquill.ai/mcp",      "headers": {        "Authorization": "${input:vaquill-authorization}"      }    }  }}

The credential goes in a prompt rather than in the file, so the file is safe to commit. VS Code asks for it on first start and remembers it. Restart the server after saving, or the tools will not appear.

Local (via uvx):

{  "servers": {    "vaquill": {      "type": "stdio",      "command": "uvx",      "args": ["vaquill-mcp"],      "env": {        "VAQUILL_API_KEY": "vq_key_your_key_here"      }    }  }}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json.

Remote:

{  "mcpServers": {    "vaquill": {      "serverUrl": "https://mcp.vaquill.ai/mcp",      "headers": {        "Authorization": "Bearer vq_key_your_key_here"      }    }  }}

Windsurf uses serverUrl for a remote server, not url, and interpolates ${env:VAR} if you would rather read the credential from the environment.

Local (via uvx):

{  "mcpServers": {    "vaquill": {      "command": "uvx",      "args": ["vaquill-mcp"],      "env": {        "VAQUILL_API_KEY": "vq_key_your_key_here"      }    }  }}

Available Tools

Tools are generated from the live Vaquill API's OpenAPI spec at startup, so the set always matches the current API. For the authoritative, up-to-date list and per-call credit costs, run the free get_pricing tool or inspect your MCP client's tool list.

The US server publishes 26 tools and the India server 22, counting search and fetch. Every one is read-only. Law change alerts (boards and watches) are deliberately not MCP tools: the API still serves them, so create and manage alerts in the Vaquill console or over the REST API directly.

The main groups (representative tools shown):

US statutes & regulations

USC, CFR, all 50 state codes, constitutions, state court rules, the Federal Register, and agency guidance.

Reading a section in context

Utility

search and fetch exist because ChatGPT's deep research and company-knowledge connectors match a corpus server by those exact names and their single-string signature, and refuse to work without them. They are thin wrappers over the typed tools above, which any client that can call them should prefer: the typed ones filter by jurisdiction, corpus, date and status, and these two do not. fetch is deliberately lenient about its id, accepting an act_id, a citation URL, a bare path, or a Bluebook citation such as 42 U.S.C. 1983.

Resources and prompts

This server is not only tools. An MCP server generated from an OpenAPI document publishes one tool per endpoint and nothing else; these two primitives are where the knowledge that is not in the API lives, and neither costs anything in the per-turn tool budget.

Resources (read, don't call):

Prompts (workflows, with the traps built in):

Indian legislation

A separate endpoint on the same host. Central and State Acts plus the instruments of the principal regulators (SEBI, RBI, MCA, IRDAI, TRAI, DGFT): 22,265 enactments and 1,098,577 individually addressable sections.

https://mcp.vaquill.ai/in/mcp        with Authorization: Bearer vq_key_your_key_herehttps://mcp.vaquill.ai/in/s/vq_key_your_key_here      key in the path, as above

Same API key, same credit balance as the US endpoint.

One endpoint serves one jurisdiction, deliberately. The US and India OpenAPI documents are disjoint, and each app derives its entire tool set from one of them, so nothing under /mcp or /s/ can expose an Indian tool and nothing under /in/ can expose a US one. The mount path selects an app; it does not filter one, so there is no per-request check to get wrong. That also keeps an integrator's context window to the jurisdiction they actually work in.

Environment Variables

Hosted server only

The variables below apply to vaquill-mcp-remote (the deployment behind mcp.vaquill.ai) and are ignored by the stdio server.

OAuth is OFF unless the whole VAQUILL_OAUTH_* group is set, and the server behaves exactly as it did before OAuth existed. A PARTIAL group raises at startup rather than serving a half-built discovery document, because Claude caches discovery globally by URL for about five minutes across all users, so a wrong one outlives the misconfiguration that produced it.

The only redirect URI registered upstream is https://mcp.vaquill.ai/auth/callback. Do NOT also register Claude's (https://claude.ai/api/mcp/auth_callback) or Claude Code's loopback: this server proxies the flow, so the upstream never sees them, and the ephemeral-port matching RFC 8252 requires is handled here.

Example Usage

Once configured, you can ask your AI assistant things like:

• "What does 17 CFR 240.10b-5 say about insider trading?"
• "Resolve 42 U.S.C. § 1983 to its section and show the full text"
• "Find California statutes on tenant repair obligations"
• "Browse the Texas statutory codes, then drill into the Penal Code"

Development

# Clone and installgit clone https://github.com/Vaquill-AI/vaquill-mcp.gitcd vaquill-mcpuv sync --all-extras
# Run locallyVAQUILL_API_KEY=vq_key_... uv run vaquill-mcp

Tests

The suite needs no API key and no network access. Every test runs against the checked-in OpenAPI fixtures in tests/fixtures/, and tests/conftest.py makes any unmocked HTTP call raise, so a clone plus uv sync is the whole setup.

uv sync --extra dev --extra remoteuv run pytest                                    # 407 testsuv run pytest -W error::DeprecationWarning       # what CI runs

CI (.github/workflows/test.yml) runs two jobs, and they are deliberately not redundant. test runs the suite against the locked environment on Python 3.10-3.13. wheel builds the package, installs it into a fresh venv without the lockfile, and runs the same suite there. The second reproduces what a customer gets from uvx vaquill-mcp: a lockfile does not constrain anyone installing the published wheel, so a dependency shipping a breaking major shows up in that job and nowhere else.

# Test with FastMCP inspectoruv run fastmcp dev src/vaquill_mcp/server.py

Security

Report a vulnerability to [email protected] (see SECURITY.md and https://www.vaquill.ai/security/vulnerability-disclosure).

Two things worth knowing before you connect this to an agent that holds other tools:

• The server is read-only. Every tool carries readOnlyHint: true and destructiveHint: false, and none writes, sends or deletes anything. The hints describe what the tools can do; they are not a security control, and a client should treat annotations from any server as untrusted unless it trusts the server.
• Returned text is third-party content. Statutes, notices, agency decisions and session laws come from government publishers (and, for some corpora, from vendor collection and OCR). We sanitise markup and restrict source links to government hosts, but we do not filter wording for instructions, and openWorldHint: false describes reach, not provenance. The server sends an instructions note saying so; clients that ignore it should still treat tool output as untrusted.

How It Works

This package is a thin MCP wrapper around the Vaquill Developer API. At startup, it fetches the OpenAPI spec from the live API and auto-generates MCP tools using FastMCP. Tool names are derived automatically from each endpoint's OpenAPI operation id, so new API endpoints show up as clean, ready-to-use tools with no package update; key descriptions are refined for optimal LLM performance. The one exception is a short exclusion list in server.py (_ROUTE_MAPS): the SSE streaming route, and the law change alert routes, which are a product decision rather than a technical limit.

Because the spec is fetched at startup (not bundled), tools automatically reflect any API changes without a package update.

Credits & Pricing

API calls consume credits. The credit costs in the tables above are current at the time of writing; the get_pricing tool and each tool's own description always reflect the live price, so treat those as authoritative if they differ.

1 credit = $0.01 USD

License

MIT

Community

Questions, ideas, or want to contribute? Join the Vaquill community on Discord.