Incident management, on-call scheduling, and intelligent analysis powered by Rootly.
Rootly MCP Server (com.rootly/mcp-server)
An MCP server that connects to the Rootly API to provide incident management, on-call scheduling, and intelligent analysis. It is designed to run with MCP clients such as Cursor, Windsurf, Claude, and other MCP-compatible applications.
๐ ๏ธ Key Features
Incident management
On-call scheduling
Intelligent analysis via Rootly API
MCP server for Rootly API integration
๐ Use Cases
Supporting incident response workflows
Managing and coordinating on-call schedules
Performing analysis related to incidents using Rootly data
โก Developer Benefits
Integrates with the Rootly API for incident-related functions
Works with common MCP clients (Cursor, Windsurf, Claude, and other MCP clients)
โ ๏ธ Limitations
Functionality details beyond incident management, on-call scheduling, and intelligent analysis are not specified in the provided source.
SSE (deprecated, removed after 2026-11-20):https://mcp.rootly.com/sse
WARNING
SSE is deprecated. The hosted https://mcp.rootly.com/sse endpoint and the self-hosted --transport sse option will be removed after 2026-11-20. Switch your client to Streamable HTTP (https://mcp.rootly.com/mcp) or Code Mode (https://mcp.rootly.com/mcp-codemode); for most clients that means changing the URL and, if your client config has one, setting the transport to http/streamable-http. SSE responses now carry Deprecation, Sunset and Link: rel="successor-version" headers.
Toolsets
Every client loads the whole tool list into the model's context: the full surface is about 125k tokens. Tools are grouped into toolsets (the same model as GitHub's and Datadog's MCP servers) so a client can load only what it needs. Recommended: ?toolsets=core, about 29 tools and 10k tokens. Without a selection you still get every tool, so existing configurations keep working.
Toolset
What it covers
core (recommended)
Everyday incident response: find and update incidents, alerts, who is on call (~29 tools)
incidents
Action items, events, roles, retrospectives, form field selections, meeting recordings
alerts
Alerts, alert sources, routes and routing rules, urgencies, heartbeats, live call routers
oncall
Schedules, rotations, shifts, overrides, escalation policies and paths
workflows
Workflows, workflow runs and playbooks; view a workflow's steps
workflow_tasks
Add or edit the steps a workflow runs: post to Slack, page on-call, open a ticket, call a URL
AI automations: prompts Rootly AI runs on its own, on a schedule or trigger
observability
Rootly Observability logs and patterns
audit
Audit log
Pick toolsets:https://mcp.rootly.com/mcp?toolsets=core,workflows, or send X-MCP-Toolsets: core,workflows
Actions that run later are separate:workflow_tasks and automations set up things Rootly does on its own afterwards (workflow steps, AI automations). Leave them out to let an assistant read and use workflows and AI investigations without being able to set those up
Everything:?toolsets=all (about 280 tools)
Read-only: add readonly=true (or X-MCP-Readonly: true), e.g. ?toolsets=all&readonly=true. Any value other than false/0/no/off turns it on. Write tools are hidden and refused, and the server also refuses any non-GET request to the Rootly API on that connection. It is a per-connection guard; for a hard guarantee, connect with a Rootly API key whose role cannot write
Deprecated: ?tool_profile= and X-Rootly-Tool-Profile. They still work (full/all mean ?toolsets=all; slim keeps its fixed ~85-tool list) and will be removed; switch to ?toolsets=. Each use is logged so remaining clients can be found
Server-wide default:ROOTLY_MCP_TOOLSETS=core,alerts sets what a client gets with no selection (default all); ROOTLY_MCP_HOSTED_TOOL_PROFILE=slim makes slim the default instead
A tool outside the selection is refused with the toolset that holds it (list_workflows is not available here: it is in the 'workflows' toolset ... Reconnect with ?toolsets=core,workflows), and the server instructions list every toolset, so an assistant can tell you the exact change. Code Mode (/mcp-codemode) always has the full catalog; readonly=true applies there too, to every tool a script calls. readonly also applies under the slim profile and ROOTLY_MCP_ENABLED_TOOLS.
If your config still points at https://mcp.rootly.com/sse, replace it with https://mcp.rootly.com/mcp before 2026-11-20 (SSE is deprecated, see above).
Agent Setup
Claude Code
With OAuth2 (recommended):
bash
claude mcp add --transport http rootly https://mcp.rootly.com/mcp
# Code Mode:
claude mcp add --transport http rootly-codemode https://mcp.rootly.com/mcp-codemode
Register the hosted server in Unity Gateway so Supervisor Agent, AI Playground, and other Databricks agents can use it:
In your workspace, open Unity Gateway > MCPs, click + MCP, and choose Connect an existing MCP server.
Pick a catalog and schema, name the MCP, and create a new connection with URL https://mcp.rootly.com/mcp.
Choose an authentication method:
Dynamic Client Registration (recommended): each user signs in to Rootly with their own account.
Bearer token: paste a Rootly API token.
Click Create & load tools, click Login if prompted, and select the tools to expose.
Or create the connection with the Databricks SDK:
python
from databricks.sdk import WorkspaceClient
from databricks.sdk.service import catalog as c
WorkspaceClient().connections.create(
name="rootly_connection",
parent="schemas/main.default",
connection_type=c.ConnectionType.HTTP,
options={
"host": "https://mcp.rootly.com",
"port": "443",
"base_path": "/mcp",
"bearer_token": "<YOUR_ROOTLY_API_TOKEN>",
},
)
The full profile exposes about 250 tools. Use Unity Gateway tool selection to expose only what your agents need. See docs/databricks-marketplace.md for the Marketplace listing details.
Rootly CLI
Standalone CLI for incidents, alerts, services, and on-call operations.
Install via Homebrew:
bash
brew install rootlyhq/tap/rootly-cli
Or via Go:
bash
go install github.com/rootlyhq/rootly-cli/cmd/rootly@latest
Choose the token type based on the access you need:
Global API Key: Full access across the Rootly instance. Best for organization-wide visibility.
Team API Key: Access limited to entities owned by that team.
Personal API Key: Access matches the user who created it.
A Global API Key is recommended for organization-wide queries and for actions that modify data, especially when workflows may span multiple teams, schedules, or incidents.
SSE (deprecated, removed after 2026-11-20) endpoint path: /sse. --transport sse logs a deprecation warning at startup; use --transport streamable-http instead.
Code Mode (experimental) endpoint path: /mcp-codemode in hosted dual-transport mode. One execute call may make at most 50 tool calls; set ROOTLY_CODE_MODE_MAX_TOOL_CALLS to a positive integer to change that.
Hosted default: full surface; ?toolsets=core for the recommended small one (see Toolsets)
Hosted slim profile (deprecated): about 85 high-usage tools via ?tool_profile=slim
Self-hosted default: full surface (ROOTLY_MCP_TOOLSETS narrows it)
To restrict either deployment to read-only tools, start the server with --no-enable-write-tools or set ROOTLY_MCP_ENABLE_WRITE_TOOLS=false.
For hosted clients, choose toolsets with ?toolsets= on the MCP URL or the X-MCP-Toolsets header (see Toolsets).
To override the hosted or self-hosted default profile entirely, set ROOTLY_MCP_ENABLED_TOOLS (or pass --enabled-tools) with a comma-separated allowlist of exact tool names. When that variable is set, it fully replaces the default selection.
To expose only a specific subset of MCP tools on a self-hosted deployment, set ROOTLY_MCP_ENABLED_TOOLS (or pass --enabled-tools) with a comma-separated allowlist of exact tool names, for example list_incidents,get_incident,get_server_version.
PostHog MCP analytics
Hosted deployments send MCP telemetry to PostHog only. Setting
POSTHOG_PROJECT_TOKEN (and optionally POSTHOG_HOST, default
https://us.i.posthog.com) sends $mcp_* events with no injected context,
conversation_id or llm_model parameters, credential scrubbing before send,
and hosted users identified by Rootly user ID only. Without the token (the
default for local and self-hosted installs) nothing is sent.
Failed calls carry the real exception type, the upstream HTTP status and an
mcp_error_expected flag. Expected failures (guided errors and Rootly API
401/403/404/409/429 answers) open no Error Tracking issue. On the stateless
hosted transport, PostHog sessions are tied together through the
Mcp-Session-Id header minted at initialize. Clients holding a cached tool
list that still send context or session_id are tolerated; the stray
argument is dropped before validation.
OpenAI Apps domain verification
Hosted deployments expose an unauthenticated domain-verification endpoint used by the OpenAI plugin directory:
Set ROOTLY_OPENAI_APPS_CHALLENGE_TOKEN to the token generated in the OpenAI submission portal. The endpoint returns that exact token as text/plain (with Cache-Control: no-store), and returns 404 when the variable is unset or empty. The token is read from the environment at request time โ never hardcode or commit it. The route is served on both the dual-transport (--transport both) and profiled streamable-HTTP hosted deployments and is not subject to Bearer authentication.
To discover the exact tool names available under your current self-hosted configuration, run:
bash
ROOTLY_API_TOKEN=<YOUR_ROOTLY_API_TOKEN> \
uv run python -m rootly_mcp_server --list-tools
This prints the effective MCP tool names after applying your current settings, including ROOTLY_MCP_ENABLE_WRITE_TOOLS and ROOTLY_MCP_ENABLED_TOOLS.
On-Call Health Integration (self-hosted only): Detects workload health risk in scheduled responders
Supported Tools
The default tool surface depends on deployment profile:
Hosted default: about 280 tools
Hosted ?toolsets=core (recommended): about 29 tools
Hosted slim profile (deprecated): about 85 tools
Self-hosted default: about 280 tools
Custom Agentic Tools
check_oncall_health_risk - self-hosted only
check_responder_availability
collect_incidents
create_incident - create a new incident with a scoped set of fields for agent workflows
create_override_recommendation
find_related_incidents
get_incident - retrieve a single incident for direct verification, including PIR-related fields
get_alert_by_short_id
get_investigation - an AI SRE investigation's verdict, evidence, causal chain and open questions (by id, or latest for an alert/incident)
get_investigation_progress - follow a running investigation; call again with after_cursor until done
get_oncall_handoff_summary
get_oncall_schedule_summary
get_oncall_shift_metrics
get_server_version
get_shift_incidents
list_endpoints
list_investigations
list_incidents
list_shifts
search_incidents
start_investigation, cancel_investigation - start or cancel an AI SRE investigation; write tools
suggest_solutions
update_incident - update native incident fields with severity validation; retrospective status changes require the Rootly UI
OpenAPI-Generated Tools
Tool naming: all tools use snake_case. The historical camelCase names
(e.g. getScheduleShifts, listIncidents) are no longer advertised in
tools/list, but remain callable as hidden aliases โ they are transparently
routed to their snake_case canonical. Update configs and ROOTLY_MCP_ENABLED_TOOLS
allowlists to the snake_case names; legacy camelCase allowlist entries are
auto-canonicalized.
Major Expansion: This version includes 50+ new endpoints covering communications, dashboards, playbooks, post-incident reviews, monitoring, and advanced form management - while carefully excluding security-sensitive operations like API key management, user creation/deletion, role management, and webhook configuration.
Delete operations remain disabled in the default tool surface.
On-Call Health Integration (Self-Hosted Only)
Integrates with On-Call Health to detect workload health risk in scheduled responders.
check_oncall_health_risk is unavailable on the hosted MCP service, including
the full and slim profiles, SSE, Streamable HTTP, and Code Mode. It is not
registered in hosted mode, so an API key or ROOTLY_MCP_ENABLED_TOOLS allowlist
cannot enable it. Ordinary on-call scheduling and incident tools remain available.
Self-hosted deployments retain this integration. Operators are responsible for
appropriate employee disclosure and data handling for their deployment.
Setup
Set the ONCALLHEALTH_API_KEY environment variable:
Finds similar historical incidents using ML-based similarity matching
Suggests solutions based on past successful resolutions
Coordinates with on-call teams across timezones
Correlates incidents with recent code changes and deployments
Creates action items and remediation plans
Provides confidence scores and time estimates
Quick Start:
bash
# Copy the skill to your projectmkdir -p .claude/skills
cp examples/skills/rootly-incident-responder.md .claude/skills/
# Then in Claude Code, invoke it:# @rootly-incident-responder analyze incident #12345
It demonstrates a full incident response workflow using Rootly tools and GitHub context.
On-Call Shift Metrics
Get on-call shift metrics for any time period, grouped by user, team, or schedule. Includes primary/secondary role tracking, shift counts, hours, and days on-call.
Complete handoff: current/next on-call + incidents during shifts.
python
# All on-call (any timezone)
get_oncall_handoff_summary(team_ids="team-1,team-2", timezone="America/Los_Angeles")
# Regional filter - only show APAC on-call during APAC business hours
get_oncall_handoff_summary(timezone="Asia/Tokyo", filter_by_region=True)
Regional filtering shows only people on-call during business hours (9am-5pm) in the specified timezone.
Returns: schedules with current_oncall, next_oncall, and shift_incidents
MCP Resources for Context
AI agents can access these resources for situational awareness:
incident://{incident_id} - Detailed incident information for specific incidents
team://{team_id} - Team details including name, color, and metadata
rootly://incidents - List of recent incidents for quick reference
rootly://oncall-status - Current on-call status across all schedules (critical for incident response)
rootly://workflow-guide - Step-by-step workflow guidance for common operations
Example usage: "Check the current on-call status" โ AI reads rootly://oncall-status resource
Shift Incidents
Incidents during a time period, with filtering by severity/status/tags.
This project was developed by Rootly AI Labs, where we're building the future of system reliability and operational excellence. As an open-source incubator, we share ideas, experiment, and rapidly prototype solutions that benefit the entire community.
Rootly AI logo