Custom MCP Servers
Servers built around the questions and actions your team needs, over SAP, NetSuite, Dynamics 365, Salesforce, HubSpot, Shopify, Paycor, TallyPrime, Zendesk and any system with an API.
MCP Server Development
MCP server development services
The Model Context Protocol (MCP) is the open standard that lets Claude, ChatGPT, Microsoft Copilot and your own agents call business systems as tools. We design, build, secure and host MCP servers over ERP, CRM, HR and data systems, starting with read-only tools and adding write actions behind approvals.
Want your systems in the hands of AI agents, safely?
Tell us which system and which questions or actions matter most. We’ll check what already exists and scope the tool surface and security model with you.
Please share the workflow and the systems involved—not credentials or customer records. We’ll agree the scope and next steps with you.
NVIDIA Inception member
Part of NVIDIA’s program for startups building with AI and accelerated computing.
About the program
Bay Area roots. Enterprise experience.
Pravin’s experience includes Google and SmartBear. Sravan previously held leadership roles at JPMorgan Chase and First Republic Bank.
Meet the foundersSan Francisco Bay Area, California
9110 Alcosta Blvd Ste H345, San Ramon, CA 94583
US-led delivery, with engineering in India. Supporting US and international organizations.Tell us where you want to start.
How we work
Check what already exists
Many vendors now ship an MCP server. We check the official server and hosted options first, and configure them when they cover the job.
Design the tool surface
Tools built around what people ask, not every API endpoint. Clear names, tight schemas, a read/write split and approvals where money or customers are involved.
Build, secure, evaluate
Implement the server, wire identity and permissions, and test it against real prompts, failed requests and edge cases before anyone in the business uses it.
Ship and hand over
Deploy as a remote MCP server, connect it to Claude, ChatGPT or Copilot Studio, and hand over documentation, runbooks and operations.
Which AI assistants can use your MCP server
One server can serve several assistants. Each client connects differently, and those differences shape hosting, authentication and network access.
| Client | Where it connects | Transport | Authentication | What to know |
|---|---|---|---|---|
| Claude (claude.ai, Claude Desktop, Cowork) | Custom connector; on Team and Enterprise an Owner adds the URL | Streamable HTTP (legacy HTTP+SSE still accepted) | OAuth 2.0, a static header credential (beta) or none | Calls come from Anthropic’s cloud, so private servers must allowlist its IP ranges. |
| Claude Code | claude mcp add, or a shared .mcp.json in the repository | stdio for local servers, Streamable HTTP for remote; SSE deprecated | OAuth 2.0 through /mcp, or headers and environment variables | Admins can fix, allow or deny servers with managed MCP settings. |
| ChatGPT (developer mode and plugins) | Developer mode, then ChatGPT Plugins; admins enable plugins by role | Streamable HTTP on a public HTTPS endpoint, or Secure MCP Tunnel | OAuth 2.1 (DCR, CIMD or a predefined client), or none for public data | On Enterprise and Edu, an admin must grant developer-mode access. |
| OpenAI API (Responses API) | An "mcp" tool with server_url, or tunnel_id for a private server | Streamable HTTP or HTTP/SSE | OAuth access token in the authorization parameter | Asks for approval before sharing data by default; allowed_tools trims the tool list. |
| Microsoft Copilot Studio | MCP onboarding wizard on the agent’s Tools page, or a Power Apps custom connector | Streamable HTTP only; SSE unsupported since August 2025 | None, API key (header or query) or OAuth 2.0 | Servers can also be registered and approved through Agent 365. |
| Custom agents (official MCP SDKs) | Your own code, using the TypeScript, Python, C#, Go, Rust or another official SDK | stdio or Streamable HTTP | OAuth 2.1-based authorization for HTTP (optional in the spec); stdio uses local credentials | You own the client, so approvals, logging and tool limits are yours to build. |
Checked against each vendor’s documentation in September 2026. Client features change often; we confirm them for your plan during discovery.
Official MCP server or a custom build?
Start with the vendor’s server when it covers your edition and your questions. Build where it stops.
| Situation | Official server | Custom server |
|---|---|---|
| Read-only questions in the vendor’s cloud edition | Usually enough; configure scope and access | Adds little |
| The system runs on-premises | Often not supported (Business Central’s server, for example, is online only) | Runs inside your network under your identity rules |
| Agents need to write, with checks | Varies; several start read-only | Validation, approval steps and an audit record on every write |
| One request spans several systems | One system per server | One tool can combine systems with consistent rules |
| The vendor has no server | Not available (Paycor and ADP, for example) | A governed server designed for the job |
| Business rules matter | Generic create, read and update tools | Encodes posting periods, tax rules or payroll logic |
The MCP server status page lists the dated, vendor-verified status of each system we integrate.
What an MCP server actually gives you
Every system your company runs, the ERP, the CRM, the ticketing tool, the warehouse, is a source of answers and a place where work happens. Today a person bridges them: reading a screen in one, typing into another. An MCP server describes the useful parts of a system as tools an AI assistant can call, so “check whether this customer is on credit hold and, if not, release the order” becomes a request rather than a procedure.
The protocol itself is simple; the value is in the design. Which tools to expose, how to name and constrain them, whose identity a call runs under, what must be logged and where a human has to approve: these are the decisions that separate a demo from something your finance team is allowed to use. That is the work we do.
Official servers changed the build-or-buy question
Over the past year many vendors released their own MCP servers, including NetSuite, Dynamics 365 Business Central, HubSpot and QuickBooks. Others have none: Paycor and ADP, for example. SAP publishes sample code for Business One rather than a supported product, and TallyPrime’s first official tools target data retrieval in Claude Desktop.
So we check what exists first. Where an official server fits, we configure and govern it. We build only what is missing: on-premises access, write actions with approvals, tools that span several systems, and the business rules generic tools cannot apply.
Systems we build MCP servers for
ERP and accounting: SAP Business One, S/4HANA and ByDesign, NetSuite, Dynamics 365 Business Central and Finance & Operations, Acumatica, Odoo, TallyPrime, QuickBooks, Xero. CRM and support: Salesforce, HubSpot, Zoho, Dynamics 365 Sales, Zendesk, Freshdesk. HR and payroll: Workday, BambooHR, ADP, Paycor. Commerce: Shopify, Adobe Commerce, BigCommerce, Amazon. Data: Snowflake, Databricks, BigQuery, PostgreSQL. And internal APIs, legacy databases and document stores nobody has an SDK for.
Why enterprises choose AIONDATA
We have shipped MCP in production: Lexicon and our agent products are MCP-native
Deep integration expertise: the same team builds ERP, CRM and eCommerce integrations, so your MCP server sits on solid data flows
Security first: identity propagation, scoped tools, audit logs and controls for personal data designed in
Model-agnostic: works with Claude, ChatGPT, Copilot Studio and custom agents through any MCP client
ISO 9001:2015 and CMMI Level 3 certified delivery processes
You own the code and can host it anywhere
Want your systems in the hands of AI agents, safely?
The enquiry form is at the top of this page. Tell us the workflow and the systems involved.
Go to the enquiry form