Mainstreet MCP

Main Street MCP

A free, industry-specific MCP server that gives a small business's AI assistants (such as Claude Desktop and Claude Code) configured facts about hours, menu/services, FAQs, staff and booking. One business.yaml file per business; no code changes required to run it. Answers depend on the accuracy of that file and the assistant using it.

The published package supports compatible local MCP clients over stdio. It is not a hosted ChatGPT app or connector. The optional HTTP mode listens on loopback and validates localhost Host/Origin headers; remote hosting, authentication and client compatibility must be configured and verified separately. A reverse proxy alone does not establish those requirements.

Quickstart (business owner)

1. Install Node.js 20+ (or use the version already on your machine — check with node -v).
2. Generate a starter config for your industry:

   npx -y mainstreet-mcp init --industry restaurant

Replace restaurant with your industry. Supported industries: restaurant, dental, law-firm, home-services, salon-spa, fitness-studio, real-estate-agent, auto-repair, insurance-agency, church.

This writes business.yaml in the current directory — edit it with your business's real name, hours, menu/services, FAQs, etc. (Use --out for a different name; every other command then needs --config .)

1. Check your edits are valid:

   npx -y mainstreet-mcp validate

A broken file (bad YAML, a typo'd industry, a field that doesn't fit) prints a specific, readable error instead of a stack trace.

1. Point your AI assistant at it. For a local MCP client (e.g. Claude Desktop) that spawns stdio servers, add an entry like:

   {

"mcpServers": { "my-business": { "command": "npx", "args": ["-y", "mainstreet-mcp", "--config", "/absolute/path/to/business.yaml"] } } }

To test the business server over local HTTP instead, run:

   npx -y mainstreet-mcp serve --http --port 3000

The HTTP server only binds to 127.0.0.1 and validates the Host/Origin headers on every request. No public endpoint, remote authentication or hosted deployment is included. A remote deployment needs a separate review of those controls and the intended client's transport requirements.

What it gives an AI assistant

Every business gets 7 universal tools regardless of industry:

• get_business_profile — name, description, contact info, address, policies
• check_open_status — open/closed right now (or at a given time), in the business's own timezone, including overnight hours and holiday/exception overrides
• search_offerings — free-text search over whatever the business sells/offers
• answer_question — looks up configured FAQs; returns found: false rather than a guess when nothing matches confidently
• list_staff — configured staff, roles, bios, credentials
• get_booking_options — how to actually book (phone/online/walk-in/email)
• submit_inquiry — lets a customer leave a message; rate-limited, sanitized, and screened by industry-specific guardrails before it's accepted

On top of that, each industry preset turns on a handful of capabilities — pluggable tool modules — matched to that industry:

Every tool ships an outputSchema and returns structuredContent, so a calling agent gets typed data back, not just prose to parse.

Guardrails

Some presets screen submit_inquiry messages before accepting them, on top of the universal sanitization (HTML/control-char stripping, length caps, rate limiting):

• dental — rejects messages containing an SSN-shaped pattern, and long messages with clinical/health detail keywords (tells the patient to call the office instead).
• law-firm — caps the message to one line / 280 characters (a topic, not case details).
• salon-spa / fitness-studio — rejects messages containing health/medical detail keywords (pregnancy, injury, medication, etc.), directing the customer to tell staff in person instead.

price_estimate outputs always carry is_estimate: true plus a disclaimer — never present as a firm quote.

Install

Claude Desktop (one-click, no terminal)

Download the latest mainstreet-mcp.mcpb from the releases page (or build it yourself — see mcpb/ below), then double-click it. Claude Desktop opens it as a Desktop Extension install prompt; it asks for one thing — a folder to keep your business.yaml in. Pick any folder you can find again; Documents is fine. No Node.js install or terminal needed.

You do not need a `business.yaml` before you install. If the folder is empty, the server starts in setup mode: ask Claude to "set up my business", and it lists the industries, writes the starter file into that folder, and interviews you to fill in your real hours, services and prices. Turn the extension off and on again when you're done and your business tools appear. The folder you picked stays the same throughout.

To build the .mcpb yourself:

npx @anthropic-ai/mcpb pack mcpb build/mainstreet-mcp.mcpb

Claude Code (plugin)

Install the bundled plugin, which registers the MCP server and adds a setup-my-business skill that interviews you and writes business.yaml for you:

claude plugin marketplace add gbrussich52/mainstreet-mcp

claude plugin install mainstreet@mainstreet

The plugin's MCP config expects business.yaml at your project root (${CLAUDE_PROJECT_DIR}/business.yaml), which is exactly what init writes — run the setup-my-business skill, or npx -y mainstreet-mcp init --industry , from that directory.

Claude Code (direct MCP add, no plugin)

claude mcp add mainstreet -- npx -y mainstreet-mcp --config /absolute/path/to/business.yaml

Generic JSON config (any MCP client)

For any client that spawns stdio servers from a JSON config (Claude Desktop's manual config, other MCP clients):

{

"mcpServers": { "my-business": { "command": "npx", "args": ["-y", "mainstreet-mcp", "--config", "/absolute/path/to/business.yaml"] } } }

Example configs

examples/ has one working .business.yaml per supported industry — the same files init copies from and the E2E test suite runs against.

Development

npm install

npm run build # tsc -> dist/ npm test # builds, then runs node --test against dist/test/*.test.js

Stack: @modelcontextprotocol/server@2.0.0 (MCP spec 2026-07-28), TypeScript strict/ESM, Zod for all schemas and config validation, yaml for parsing business.yaml. All dependencies are pinned to exact versions (no ^).

Project layout

• src/config/schema.ts — the business.yaml Zod schema (BusinessConfigSchema) and types.
• src/config/load.ts — loads + validates a config file, including per-capability config and the industry/capability compatibility check.
• src/hours.ts — timezone-aware open/closed logic (Intl-based), overnight ranges, date exceptions, forward-scanning for the next open/close time.
• src/search.ts — the recall-first free-text search used by search_offerings and answer_question.
• src/inquiries.ts — InquirySink: sanitization, rate limiting, guardrail hook, and file/webhook delivery for submit_inquiry.
• src/capabilities/* — the 9 pluggable capability modules (config schema + tool registration each).
• src/presets/* — the 10 industry presets: which capabilities are on, the instructions text injected into the server, and any guardHooks.
• src/tools/universal.ts — the 7 universal tools every business gets.
• src/server.ts — createServer(config): builds the McpServer for a loaded config.
• src/cli.ts — mainstreet-mcp entry point (stdio default, init, validate, serve --http).
• src/test/*.test.js (compiled from .ts) — unit tests (schema, search, hours, inquiries/ guardrails) plus an E2E suite that spawns the built stdio server for every example config and drives it with the real MCP client SDK.

Testing notes

• npm test runs both unit tests and the E2E suite (src/test/e2e.test.ts), which spawns dist/cli.js as a child process via StdioClientTransport for every file in examples/, calls tools/list, then calls every registered tool and asserts it returns structuredContent without isError.
• serve --http binds a local port, which some sandboxed shells block by default; test it outside such a sandbox if EPERM/listen errors appear.

The submit_inquiry tool advertises openWorldHint: true when configured for webhook delivery and false for local file delivery. It remains a non-idempotent write tool in both modes.

Status

Presets, capabilities, universal tools, CLI, and all 10 example configs are built and verified (build, unit tests, E2E stdio tests, CLI validate success/failure paths, and an HTTP smoke test all pass — see the build's final report for exact output). Version 0.1.4 is published on npm, in the GitHub release and in the official MCP registry as io.github.gbrussich52/mainstreet-mcp. These are package distribution channels; they do not provide a hosted endpoint or prove a particular customer's integration. The desktop bundle opens business setup; workflow kits use the CLI configuration below.

Workflow kits: caller-supplied facts to a review packet

Version 0.1.4 adds three stateless kits. Each exposes exactly two tools: analyze_workflow and prepare_workflow_draft. They evaluate supplied structured facts with explicit sample policies and prepare an unsent review draft when inputs are adequate. No provider account, live source verification, storage, messages, money movement or fulfillment changes are included. Your MCP host may send tool inputs to its model provider; use fictional inputs first and follow your organization's data rules.

npx -y mainstreet-mcp@0.1.4 kit quote-followup

npx -y mainstreet-mcp@0.1.4 kit invoice-discrepancies npx -y mainstreet-mcp@0.1.4 kit order-handoff

Use one kit per configured MCP server. Try the fictional demos and request a scoped implementation. These free kits do not include connected customer systems. Check native CRM, accounting or order software before commissioning a custom integration.

{ "mcpServers": { "quote-review": { "command": "npx", "args": ["-y", "mainstreet-mcp@0.1.4", "kit", "quote-followup"] } } }

Ask: “Review this quote using the supplied timestamps and follow-up policy. Prepare a draft only if the tool says it is ready.” Send all required fields from the published tool schema, using null fo