Skip to main content
AI Interoperability

Published March 2026

Building 608 MCP Tools: AI-to-AI Communication for Business Operations

Most MCP implementations have 10-50 tools. We built 608 across 14 servers — and every customer gets their own auto-generated MCP endpoints. Here's how we made AI-to-AI communication work at scale.

Why 608 Tools?

The number isn't a vanity metric. It's a consequence of building a platform that covers every aspect of business operations:

  • 429 database tables each need CRUD operations exposed to AI agents
  • 52 industries each have domain-specific tools (healthcare has SOAP notes, e-commerce has cart management)
  • 14 AI agents each have specialized tools for their domain
  • 130 Google Workspace tools alone (Gmail, Calendar, Drive, Sheets, Docs, Tasks, Forms, Chat, Meet)
  • Business logic tools that go beyond CRUD: price optimization, lead scoring, campaign forecasting, inventory redistribution

When you build full business infrastructure — CRM, payments, POS, inventory, booking, marketing, content, analytics — and give AI agents the ability to operate all of it, you end up with hundreds of tools. Not by choice, but by necessity.

Two-Tier Architecture: Internal + External

Not all 608 tools are public. The system has two tiers with different security models:

AspectInternal MCPExternal MCP
Tools~600~19 public endpoints
Users14 internal AI agentsExternal AIs (ChatGPT, Claude, Perplexity)
AuthJWT + API key + scopesNo auth (read-only) or optional API key
DataFull database accessSanitized public data only
Write accessYes (create, update, delete)No (read-only)
Tenant isolationcompany_id from JWTcompany_id from domain detection

Internal MCP is how Sarah books an appointment, Marcus runs a marketing analysis, and ADA orchestrates multi-agent workflows. External MCP is how ChatGPT answers “what are Best Dental's hours?” by querying the business's public MCP endpoint.

The 14 Server Categories

Tools are organized into logical servers by business domain:

CRM & Contacts

60+

Contact management, deals, pipelines, tasks, communication templates

Commerce Operations

80+

Pricing, tax, inventory, orders, payments, rewards, commissions

Marketing & Content

70+

Email campaigns, drip sequences, A/B testing, ROI analysis, targeting

Google Workspace

130

Gmail, Calendar, Drive, Sheets, Docs, Tasks, Forms, Chat, Meet

Knowledge Base & AI

30+

KB search, generation, vector search, context export, 3-layer retrieval

E-Commerce

40+

Products, storefronts, shopping cart, reviews, payment QR codes

AI Orchestration

30+

ADA coordination, mission dispatch, knowledge graph traversal

Development Tools

25+

Route introspection, test running, migrations, CI/CD

Platform Monitoring

25+

Performance metrics, resource usage, security events, self-healing

Super Admin

20+

Tenant management, error monitoring, platform operations

Healthcare

15+

Medical case management, SOAP notes, patient records

Content Creation

30+

Agent-specific tools for Marcus (marketing), Maya (brand), Ace (web)

Landing Pages

15+

Page management, analytics, template system

Integration & Data

20+

Webhooks, data mapping, CSV analysis, OAuth

Every Customer Gets Their Own MCP

This is the feature that makes Solid#'s MCP implementation unique: every business on the platform automatically gets auto-generated, domain-aware MCP endpoints.

When a plumber signs up and gets acme-plumbing.solidnumber.com, the following endpoints are immediately live:

# Auto-generated for every company
GET /.well-known/mcp.json # MCP discovery manifest
GET /.well-known/ai-plugin.json # ChatGPT/Claude plugin
GET /api/mcp # Full business profile
GET /api/mcp/openapi.json # OpenAPI spec
GET /llms.txt # LLM-readable summary
GET /sitemap.xml # Sitemap
GET /robots.txt # Crawler directives

The manifest content is generated dynamically from the company's actual data — their services, hours, FAQ, blog posts, landing pages. When an AI assistant is asked about that business, it can discover and query these endpoints automatically.

3-Layer Domain Detection

How does the system know which company a request is for?

  1. X-Company-ID header — explicit override (internal use)
  2. Custom domains table — verified domains like bestdental.com
  3. Subdomain mappings*.solidnumber.com subdomains

An external AI visiting bestdental.com/.well-known/mcp.json automatically gets Best Dental's manifest. No configuration needed. No API key. The domain is the identity.

Security: 33 Scopes, Audit Trail, Rate Limits

608 tools with full database access is a security challenge. The MCP server implements:

  • 33 granular scopes. Each API key has specific permissions:kb:read, crm:contacts:write, ops:payments:charge. A key with kb:read can search the knowledge base but can't modify contacts or process payments.
  • Mandatory tenant isolation. Every tool receives company_id injected by middleware — not from the request. Tools don't trust caller-provided company IDs.
  • Complete audit trail. Every MCP operation is logged to MCPAuditLogwith the tool name, company, user, request, response, and timing.
  • Per-key rate limiting. Different rate limits per key, per IP, per endpoint. Public discovery endpoints: 120 req/min. Authenticated operations: configurable per key.

Multi-Agent Orchestration

The 608 tools aren't used by a single AI. They're shared across 14 agents, each with different access patterns:

  • Sarah (customer service) uses CRM contact tools, booking tools, and KB search to handle inbound inquiries
  • Marcus (growth) uses marketing analytics, campaign tools, and competitor analysis to drive growth
  • ADA (orchestrator) uses mission dispatch and coordination tools to route work across agents
  • Devon (operations) uses monitoring, self-healing, and deployment tools

When a customer asks Sarah about pricing and Sarah doesn't have the answer, she can use KB search tools to find it. If a complex question needs analytics, ADA can dispatch it to Marcus with the relevant context. The tools are the shared language between agents.

Not Just CRUD: Intelligent Tools

Most MCP servers expose database CRUD operations. Ours do too — but the most valuable tools are the ones that embed business logic:

  • price.suggest_optimal_price — analyzes competitor pricing, demand patterns, and margin targets to recommend pricing
  • crm.deals.forecast — predicts deal closure probability based on pipeline stage, activity history, and industry benchmarks
  • marketing.analytics.suggest_improvements — analyzes campaign performance and recommends specific changes
  • kb.vector_search — semantic search across the knowledge base using embeddings, not just keyword matching
  • ecommerce.abandoned_cart.recover — triggers intelligent follow-up sequences for abandoned carts

These tools let AI agents make decisions, not just read data. The difference between “here are your sales numbers” and “your conversion rate dropped 15% — here's why and here's what to do about it.”

What We Learned

  1. Organization is everything at scale. 608 tools in a flat list is unusable. 14 logical server categories with clear naming conventions (crm.contacts.search, ops.inventory.check) make tools discoverable by AI agents without a manual.
  2. Per-tenant MCP is a competitive moat. Most AI platforms have one MCP endpoint for the whole platform. Giving every customer their own auto-generated MCP means external AIs can discover and recommend individual businesses — not just the platform.
  3. Domain detection replaces API keys for discovery. Public MCP endpoints don't need authentication because the domain itself identifies the tenant. This makes discovery frictionless for external AI systems.
  4. Scopes prevent tool sprawl from becoming a security risk. 608 tools with full access is dangerous. 608 tools where each key can only access its authorized subset is manageable. Scopes make large tool counts safe.
  5. Business logic tools are 10x more valuable than CRUD. An AI agent that can read inventory is useful. An AI agent that can predict when you'll run out of a product and auto-generate a purchase order is transformative.

AI-to-AI Is the Future of Business Software

MCP isn't just a protocol for tool calling. It's how AI systems will discover and interact with businesses. When a customer asks ChatGPT “find me a plumber near me,” the plumber whose website has an MCP endpoint with real-time availability and pricing has an advantage over the one with a static HTML page.

Every Solid# business is AI-discoverable from day one. 608 tools. Per-tenant endpoints. Domain-aware routing. This isn't a roadmap item — it's live in production.

Building 608 MCP Tools: AI-to-AI Communication for Business Operations | Solid# Research | SolidNumber