PagerDuty MCP server: hosted, 18 tools, OAuth or API key

PagerDuty moved its MCP server to a hosted endpoint and archived the local one. The credential now decides what an agent can change.

Checked 6 October 2026 · from PagerDuty's MCP Server docs, its archived repository and the local server's tools/list response

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

What is the PagerDuty MCP server?

It is PagerDuty's server for the Model Context Protocol. The agent gains tools for reading incidents, schedules, escalation policies and services.

PagerDuty runs it as a hosted service. Each request carries an OAuth token or an API key, and the agent acts with that credential's permissions. Some filters, such as scoping incidents by team or assignee, need user-level credentials.

Each hosted tool takes an action. browse_incidents covers list, get, list_alerts, list_notes and five more, so one tool stands in for several API calls.

The deprecated local server ran on your machine with a user API key. It started read-only: its write tools were absent from the tool list until it was started with --enable-write-tools.

PagerDuty covers production observability partially through incident response, and reaches nothing in code review, release, agent observability or cost.

For an agent, PagerDuty answers questions about who was paged and what happened next, and nothing about why the system broke. An agent connected here can tell you which service alerted, who was on call and how long acknowledgement took. It cannot show you the failing query behind it, so this connector sits beside a telemetry one.

What can an AI agent do with PagerDuty?

Each hosted tool takes an action, such as list, get or add_note, so one tool covers several API calls. The manage_ tools are always listed on the hosted server, and a Scoped OAuth client decides which of their calls succeed.

What the tool definitions cost in context

A client that loads every tool definition at the start of a session spends this much context before the first question. Claude Code defers definitions by default through tool search, and loads a tool's definition when the agent needs it. The method below gives the version and date measured.

A context window is the amount of text a model can hold at once. A client that loads every tool definition upfront sends each tool's name, description and input schema to the model when the session starts. That list is charged whether or not a single tool is called.

These figures were measured on the deprecated local server, on 19 August 2026. Enabling writes there was expensive twice over. It added 38 tools that can change your on-call configuration, and it took the connector from roughly 6% of a 200,000-token window to roughly 16%.

The hosted server's 18 tools carry different definitions, and they have not been measured.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted 19 August 2026 from tools/list on pagerduty-mcp 1.1.0, run locally with a placeholder key. It ran once with defaults and once with --enable-write-tools.

Registration is static, so no PagerDuty account was involved. Figures cover name, description and input schema. The two models in column one share the o200k_base encoding, so they get one count. The spread is formatting, about 52%. Reasoning tokens are output, spent per request. The tool list is billed as input whenever the client loads it.

The Claude figures are measured rather than encoded. Anthropic publishes no local tokenizer, so the definitions were sent to the API's count_tokens endpoint and a no-tools baseline subtracted. The range covers two tokenizer generations rather than two serializations. Haiku 4.5 and Opus 4.6 sit at the low end, Opus 5 and Sonnet 5 at the high end. The gap between them is about a third.

PagerDuty MCP server limits: where answers come back incomplete

PagerDuty documents every limit below. The first comes from the hosted server's own limitations, and the second from the deprecated local server's README.

How to configure PagerDuty MCP for agents

1. 1

   Use the hosted server

   PagerDuty deprecated the local server on 4 September 2026 and points users to mcp.pagerduty.com/mcp.

2. 2

   Scope through the credential

   Tool filtering is unavailable on the hosted server, so all 18 tools are listed. A Scoped OAuth client limits which calls succeed.

3. 3

   Prefer user-level credentials for incident work

   PagerDuty notes that some filters, such as scoping incidents by team or assignee, need user-level authentication.

4. 4

   Create the OAuth client first

   PagerDuty does not support Dynamic Client Registration, so OAuth needs a client ID and secret from your account.

5. 5

   If you keep the local server, keep its read-only default

   Its write tools were omitted from the tool list until --enable-write-tools was set. On 19 August 2026, that default cost about 12,695 tokens, and enabling writes took it to about 31,904.

Paste this into your agent’s instructions

PagerDuty's silent failure is absence rather than truncation: change events that were never sent, and calls the credential is not scoped for. The block makes the agent distinguish an empty result from an unconfigured integration.

When you answer using PagerDuty MCP tools, state:
- Whether the credential allows changes. On the hosted
  server, manage_ tools are always listed, and the OAuth
  scope decides which calls succeed.
- The time window you searched, and the service or team scope.
- Whether change events exist for the period. They appear only
  if a pipeline sends them, so an empty result is not evidence
  that nothing shipped.
When a search returns nothing, say so, and state the
constraints above.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Sometimes. PagerDuty's REST API is well documented, and a coding agent can call it with the same API key the hosted server accepts.

The server earns its place through its tool descriptions and grouped actions. An agent choosing endpoints without descriptions chooses less well.

The trade is context. The hosted server's 18 definitions were not measured here. On the deprecated local server, 63 read-only definitions cost roughly 6% of a 200,000-token window in a client that loads every definition upfront.

Connect the server for open-ended incident investigation. Call the API directly when the question is fixed and you want the context back.

PagerDuty MCP server setup

The server is hosted at https://mcp.pagerduty.com/mcp. EU accounts use https://mcp.eu.pagerduty.com/mcp.

An API key goes in the header as Authorization: Token token=<key>. OAuth takes a client ID and secret from your PagerDuty account, or a token sent as Authorization: Bearer <token>.

The local server is deprecated, and its repository has been read-only since 4 September 2026. Its setup appears at the end for teams that still run it.

PAGERDUTY_API_KEY=<your-key> claude mcp add-json pagerduty \
  '{"type":"http","url":"https://mcp.pagerduty.com/mcp",
    "headers":{"Authorization":"Token token=${PAGERDUTY_API_KEY}"}}'

MCP_CLIENT_SECRET=<client-secret> claude mcp add-json pagerduty \
  '{"type":"http","url":"https://mcp.pagerduty.com/mcp",
    "oauth":{"clientId":"<client-id>","callbackPort":8080}}' \
  --client-secret

{"mcpServers":{"pagerduty":{
  "url":"https://mcp.pagerduty.com/mcp",
  "headers":{"Authorization":"Token token=<your-key>"}}}}

[mcp_servers.pagerduty]
url = "https://mcp.pagerduty.com/mcp"
http_headers = { Authorization = "Token token=<your-key>" }

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

{"mcpServers":{"pagerduty":{"command":"npx",
  "args":["-y","mcp-remote","https://mcp.pagerduty.com/mcp",
    "--header","Authorization:${AUTH_HEADER}"],
  "env":{"AUTH_HEADER":"Token token=<your-key>"}}}}

{"mcpServers":{"pagerduty":{
  "httpUrl":"https://mcp.pagerduty.com/mcp",
  "headers":{"Authorization":"Token token=<your-key>"}}}}

{"inputs":[{"type":"promptString","id":"pagerduty-api-key",
  "description":"PagerDuty API Key","password":true}],
 "servers":{"pagerduty":{"type":"http",
  "url":"https://mcp.pagerduty.com/mcp",
  "headers":{"Authorization":"Token token=${input:pagerduty-api-key}"}}}}

{"mcpServers":{"pagerduty":{"type":"http",
  "url":"https://mcp.pagerduty.com/mcp",
  "headers":{"Authorization":"Token token=<your-key>"}}}}

claude mcp add-json "pagerduty" '{"command":"uvx",
  "args":["pagerduty-mcp"],
  "env":{"PAGERDUTY_USER_API_KEY":"<your-key>"}}'

uvx pagerduty-mcp --enable-write-tools
# env: PAGERDUTY_USER_API_KEY=<your-key>

docker run -i --rm -e PAGERDUTY_USER_API_KEY \
  pagerduty-mcp:latest

Read-only access, permissions and security

• Write tools are listed by default on the hosted server

  PagerDuty states that tool filtering is unavailable there. A Scoped OAuth client limits which tools succeed, so scope the client before connecting an agent.

• The hosted server authenticates every request

  It accepts an OAuth token or a PagerDuty API key in the Authorization header. A team on OAuth avoids placing a user API key in a local process on every machine.

• User-level and account-level credentials both work

  A user credential carries one person's permissions. An account-level key or app token reaches further, so give the agent the narrowest credential that answers its questions.

• The deprecated local server was weaker on HTTP

  PagerDuty documented that its streamable-http and sse transports, which is server-sent events, had no authentication. Every request used the single API key the server held.

• The local server declared its own read-only tools

  63 of its 101 carried a readOnlyHint of true, matching its default set exactly, as counted on 19 August 2026.

Troubleshooting

A `manage_` tool call is refused

The credential lacks that permission. On the hosted server, every tool is listed, and the scope of the OAuth client decides which calls succeed.

401 on connect

The Authorization header is missing or malformed. An API key goes as Token token=<key>, and an OAuth token as Bearer <token>. EU accounts use mcp.eu.pagerduty.com/mcp.

OAuth fails at client registration

PagerDuty does not support Dynamic Client Registration. Create an OAuth client in your PagerDuty account, then give the MCP client its ID and secret.

Incident filters by team or assignee fail

PagerDuty notes that some tools and filters need user-level authentication. An account-level credential cannot scope incidents to the person behind it.

Change events are missing for the period

Change events exist only if your deployment pipeline sends them. An unintegrated pipeline produces the same empty result as a quiet period.

Another machine is using your local server

The deprecated local server's HTTP transports have no authentication. Move to the hosted server, or bind the local one to localhost behind an authenticating proxy.

PagerDuty MCP server: Reference

What engineers report

PagerDuty archived the local server's repository on 4 September 2026, at Apache 2.0 and 79 stars. Public discussion is small, so the two issues below are the whole record rather than a sample.

“The PagerDuty MCP server is great! But it exposes a bit too many tools, especially given LLMs today will have degraded performance with the increase number of MCP tools available to them.”

“The MCP implementation does not support OAuth2 Tokens. The underlying api expects the Authorization header to be slightly different between API keys and OAuth2 Tokens.”

Repository facts from the GitHub API, retrieved 6 October 2026. Hosted tool names from PagerDuty's MCP Server page, updated 4 September 2026. Local tool counts come from the local server's own tools/list response on 19 August 2026, captured once with defaults and once with --enable-write-tools.

Should you connect an agent to PagerDuty?

Yes. Best for PagerDuty teams on the hosted server, with a read-scoped credential for investigation work.

Strongest fit

Teams who want an agent reading incident history, on-call schedules and escalation policies during or after an outage.

Main advantage

Hosted, with OAuth or an API key, so nothing runs locally and every request is authenticated.

Main weakness

Tool filtering is unavailable on the hosted server, so the write tools are always listed.

Operational risk

An agent with a broad credential can create incidents and change schedules. Scope the OAuth client before connecting it.

Frequently asked questions