MCP CRM: The Customer Relationship Manager Built for the Model Context Protocol

What MCP is, in one paragraph

Model Context Protocol is an open standard introduced by Anthropic in late 2024 that lets AI models connect to external tools and data. It’s been called “USB-C for AI,” which is the right metaphor: a single protocol that any compliant client (Claude Desktop, Claude Code, Cursor, custom agents) can use to access any compliant server (CRMs, databases, file systems, APIs). Before MCP, every integration was a one-off. After MCP, you write the server once and every AI client can use it.

The Linux Foundation took stewardship of MCP in 2025. OpenAI, Microsoft Copilot, Google’s Gemini ecosystem, and most major AI vendors have adopted or are adopting the spec. It’s not a proprietary Anthropic thing anymore. It’s the standard.

Why MCP matters for CRMs specifically

CRMs are the highest-leverage place to plug an AI agent in. They hold customer data, deal data, activity data, and (for any serious sales motion) the pipeline of work that drives revenue. An AI agent with read and write access to your CRM can do the work that used to take 30% of your team’s time: drafting outreach, scoring leads, updating records, building reports, running automations.

The problem is that getting AI agents into CRMs has historically been painful. Three options existed:

1. Use the CRM vendor’s proprietary AI. Salesforce Einstein, HubSpot Breeze, Zoho Zia. They work, but they’re closed systems. You can’t swap in Claude when you need better writing or GPT when you need long-context analysis. You’re locked to whatever model the vendor picked.

2. Build a custom REST integration. Possible, but you’re writing tool definitions, handling auth, managing rate limits, and updating your integration every time the CRM’s API changes. It’s a project.

3. Wait for a third-party plugin. Slow and fragmented. Zapier and n8n help, but they’re not designed for AI agent workflows; they’re designed for trigger-action automation.

MCP collapses all three options into one: the CRM publishes an MCP server, and any MCP-compatible agent can use it. No proprietary lock-in. No custom integration code. No third-party plugin tax.

This is why we built Conduyt the way we did.

The Conduyt MCP server

Conduyt’s MCP server exposes 104 tools across 20 modules. Every major CRM operation has at least one tool: contacts, deals, pipelines, tasks, companies, activities, tags, users, automations, email sequences, messaging, calendar, forms, invoices, webhooks, drip campaigns, lead scoring, products/templates, smart lists/custom fields, bulk operations, and AI/discovery utilities.

A few specifics that matter:

One consistent naming convention. Every tool is conduyt_<module>_<action>. conduyt_deals_list. conduyt_contacts_create. conduyt_email_sequences_enroll. The AI doesn’t need to learn a weird taxonomy. It learns the pattern once and predicts the rest.

Read and write coverage. This is where most CRM MCP servers come up short. HubSpot’s Claude connector, as of early 2026, is read-only. Many community-built MCP servers for Salesforce and Zoho are read-only too. Conduyt’s MCP server supports full CRUD on every major object. Your AI agent can actually do work, not just answer questions.

Schema discovery built in. The conduyt_api_catalog tool returns the full catalog of every endpoint, parameter, and return type. Your AI agent calls it first, learns the surface, and operates against the live schema. No SDK to update when we add new tools. No version drift.

Safety primitives at the protocol layer. Scoped API keys, dry-run mode, sandbox mode, confirmation tokens for destructive actions, action budgets, full audit trail. The MCP server respects all of these. Your AI gets exactly the access you grant it, no more.

Native Claude Desktop and Claude Code support. Drop a JSON snippet into your Claude config file, restart, and the 104 tools are live. We tested first on Claude because Anthropic wrote MCP, but the protocol is open and Cursor, Continue, and any custom MCP client work the same way.

How Conduyt’s MCP compares

A short comparison of where MCP support stands across major CRMs in 2026:

Conduyt

Native MCP server. 104 tools, 20 modules. Full read/write. Built in, no add-on, no premium tier. $299/mo flat covers it.

HubSpot

MCP server available. Read-only for the Claude connector specifically. The “remote MCP server” (developer-facing) supports more operations but requires the right plan tier. Access is gated by HubSpot’s pricing structure.

Salesforce

MCP support announced and rolling out via Agentforce. Premium-gated. Available on higher tiers. Limited tool count compared to what’s available through the underlying Salesforce API.

Zoho

Zoho MCP available across multiple Zoho apps (CRM, Mail, Calendar, etc.). OAuth-based. Read and write supported. Strong if you’re already a Zoho One customer; less compelling if you’re picking a CRM fresh.

Pipedrive

No native MCP server as of early 2026. Community-built MCP servers exist but aren’t officially supported.

Close, Salesflare, Freshsales, Capsule, etc.

No native MCP support. Community or third-party MCP servers may exist but aren’t first-party.

Open-source / self-hosted

Relaticle (AGPL-3.0) is the most prominent open-source MCP CRM. 30 tools. Self-hosted, full data ownership. The right pick if you want to own everything end-to-end and have the engineering capacity to maintain it.

The honest read: MCP support is becoming the dividing line between CRMs built for AI agents and CRMs that bolted AI features onto a 2010-era architecture. Conduyt and Zoho are leading on first-party MCP. HubSpot and Salesforce are catching up but gating it behind plan tiers. Most other CRMs aren’t there yet.

What you can actually do with an MCP-connected CRM

A few concrete examples of what “104 tools, any MCP client” looks like in practice:

Natural-language pipeline queries. Ask Claude Desktop “what’s the status of my top 10 deals?” and Claude calls conduyt_deals_list, ranks by value, returns a real answer. No dashboard navigation. No exporting to spreadsheets.

Bulk operations from a single sentence. “Reassign all deals stuck in proposal for more than 14 days to Maya.” Claude calls conduyt_deals_list with stage and age filters, then conduyt_deals_bulk_assign. Three seconds, one sentence, twelve deals moved.

Multi-step workflows the AI runs end-to-end. “Find leads who haven’t responded in 30 days, draft a re-engagement email, enroll them in the warm-leads drip campaign.” Claude calls four or five tools in sequence, you approve the drafts, the campaign launches.

Audit-friendly automation. Every tool call is logged. If your AI does something weird, you can see exactly what it called, when, and with what parameters. Compliance teams stop blocking AI integrations because the visibility is built in.

Mixed-agent workflows. Claude handles writing tasks, GPT runs analysis on the data Claude pulled, a custom agent triggers downstream actions in your other tools. All of them speak MCP. All of them work with the same Conduyt server.

This isn’t theoretical. The Bring Your Own AI Agent guide on this site walks through specific setups using exactly this pattern.

Setting up the Conduyt MCP server

For Claude Desktop or Claude Code users, the entire setup is a JSON snippet you drop into your config file:

{
  "mcpServers": {
    "conduyt": {
      "command": "npx",
      "args": ["tsx", "src/index.ts"],
      "cwd": "~/path/to/conduyt-mcp",
      "env": {
        "CONDUYT_API_URL": "https://conduyt.app",
        "CONDUYT_API_KEY": "cdy_your_key_here"
      }
    }
  }
}

Restart your client. The 104 tools are live.

For Cursor, Continue, or custom MCP clients, the config format varies by client but the server endpoint is the same. The full setup guide lives in the developer docs.

For non-MCP integrations (Zapier, n8n, custom REST), the same operations are available through Conduyt’s REST API. Same auth, same permissions, same audit log. MCP isn’t a wall; it’s an option.