ServiceNow MCP Server: Connect Claude or ChatGPT to ServiceNow (2026)

ServiceNow has its own MCP server: the Model Context Protocol Server plugin, managed from the MCP Server Console, lets an admin publish chosen tools that Claude, ChatGPT or any other MCP client can call with OAuth. On 28 October 2026 ServiceNow and OpenAI also put a ServiceNow MCP plugin in OpenAI's marketplace, so a ChatGPT user can find similar incidents, check change risk and update records without leaving the chat. That gives you three ways in: ServiceNow's native server, the packaged ChatGPT plugin, or an open-source server. Below: what each needs, the steps for Claude, the settings that break it, and when you need a server of your own.

---

The three options

Option 1: ServiceNow's native MCP server

ServiceNow's MCP Server Console documentation describes it as giving "secure and governed access to functionality on a ServiceNow instance for AI applications". An admin creates a server and chooses the tools and inputs it exposes; nothing else on the instance is reachable through it. Feature access depends on your license, so confirm with your ServiceNow account team before planning around it.

ServiceNow's developer advocate published a full walkthrough connecting Claude in August 2026. The short version:

1. Check the release and plugins. The walkthrough needs Australia Patch 2 or later. Confirm Now Assist Admin Console (sn_nowassist_admin) is present, then install Model Context Protocol Server (sn_mcp_server); the author warns it "may not install cleanly" without the first.
2. Build what the tools call. In the walkthrough, a scoped app with a Scripted REST API: a GET to read an incident and a POST to update one. Give the POST an OpenAPI schema, or the tool has no body inputs. Flow actions, subflows and Now Assist skills also work as backends.
3. Create the tools in the MCP Server Console. Each gets a label and a description, and the description is what the model reads to decide when to call it, so write it like an instruction: what it does, and when to use it.
4. Create and activate the server with at least one tool.
5. Register an OAuth client (application registry) with Claude's redirect URL. Set the token format to JWT: on the default, Opaque, "the authentication fails at connection time".
6. Add a custom connector in Claude with the server URL, client ID and client secret, sign in, and approve the consent screen. Adding custom connectors is covered in how to add a custom connector in Claude.

The scope to tighten before production

The walkthrough uses the useraccount OAuth scope because it is fast for a lab, and says plainly that it "grants the client broad access to every REST API in the instance under the authenticated user's permissions". The MCP server only exposes the tools you published, but the token itself is wide. For anything beyond a test instance, create a custom scope limited to what the tools need.

Where the data goes

ServiceNow's documentation says data may move from your instance to a centralized ServiceNow environment, possibly in another data-center region and on a third-party cloud such as Microsoft Azure, and that ServiceNow uses inputs, outputs and edits to improve its products unless you opt out in the Now Assist Admin console. Read that section with whoever owns data protection, and decide on the opt-out before rollout, not after.

Option 2: the ChatGPT plugin

The ServiceNow and OpenAI announcement of 28 October 2026 says "the ServiceNow MCP plugin is available in the OpenAI marketplace today". It is powered by ServiceNow's Action Fabric, and its example is an IT agent who finds similar incidents, assesses change risk and updates the record from ChatGPT, while "ServiceNow remains the system where those actions are executed and governed".

What is dated and what is not:

• Available now: the plugin in OpenAI's marketplace.
• By the end of 2026: enhanced plugin capabilities.
• Preview, no date: richer interactive experiences inside ChatGPT, such as reviewing alerts and updating an incident.
• Q1 2027: voice in ServiceNow's Control Tower. January 2027: early access to OpenAI-powered security capabilities, general availability in Q2 2027.

The announcement does not say which ServiceNow licenses the plugin needs, so ask before you budget. If your company uses ChatGPT Business or Enterprise, this is the shortest path. If it uses Claude, or both, the native server above serves any client.

Option 3: the open-source server

echelon-ai-labs/servicenow-mcp is a Python server (3.11 or later) under the MIT license, with about 75 tools across incidents, the service catalog, change management, agile work, knowledge bases, users and groups, and script includes. It runs over stdio for local clients or as an SSE server, signs in with basic auth, OAuth or an API key, and lets you load only a package of tools rather than all of them.

It is a good way for a developer to see what an assistant can do with ServiceNow in an afternoon. It is not how we would roll ServiceNow out to a team: the credentials sit in a config file on someone's machine, every call runs as that one account, and nobody at ServiceNow supports it. Its SSE transport also predates Streamable HTTP, which is what remote MCP clients use now (remote MCP servers explained).

When none of them is enough

ServiceNow's own server is the right default. It stops being enough when:

• One action crosses systems. Closing an incident should also credit a customer in your billing system, or a change request should check stock in your ERP. ServiceNow's server reaches ServiceNow.
• Writes need a safety net. Bulk updates or reassignments should show the person exactly what will change and wait for a confirmation, not just execute (MCP elicitation).
• Your instance is behind. If you are not on a release that ships the MCP plugin, a server in front of ServiceNow's REST APIs gets you there without an upgrade project.
• Data must stay put. If ServiceNow's centralized processing doesn't fit your data rules, the assistant can call a server you host, in your region, with the model of your choice.

What we do

We build and run MCP servers in production. Some live inside the application they serve: our CRM answers on its own /mcp endpoint, and so does a time-tracking and project-profitability dashboard we built for a client. Others stand in front of systems we don't own: Google Sheets, Linear, spreadsheet files, a 1.3-million-row wine database. They run stateless on the official Go SDK, users sign in with OAuth as themselves, each token's scope comes from the tools themselves (read-only tools need read, everything else needs write), and tenancy is enforced in the server's code. Details in remote MCP servers and MCP security.

What we haven't done yet: a production server against ServiceNow. The pattern is the same as our Salesforce MCP work, but if you are the first ServiceNow project, we'll say so on the call, and scoping is where we confirm what your instance exposes. Our published prices: scoping one workflow costs $2,300 to $4,700, a first workflow in production starts at $17,500, and running it starts at about $470 a month (what an MCP server costs).

Related: how to connect Salesforce to Claude with MCP, who can build a custom MCP server, MCP server examples, and what MCP is.