Skip to main content
Out-of-the-box integrations cover the popular tools: Slack, Linear, PagerDuty, Notion, Teams, GitHub, GitLab. For everything else you run internally, RubixKube supports two custom paths. This tutorial walks through both.

​
Two paths, different trade-offs

Custom MCP Server

Run a Model Context Protocol server that exposes your tool. Best for rich, bidirectional integrations where the agent needs tool-shaped access.

Custom REST Integration

Point RubixKube at a REST endpoint with an OpenAPI spec. Best for simple read-or-write integrations with an existing API.
If the tool you want to integrate already ships an MCP server, use that. If it only has a REST API, the REST path is simpler. If neither exists, write an MCP server for anything beyond a single endpoint.

​
Prerequisites

  • Admin access to your RubixKube workspace.
  • The internal tool’s API reachable from RubixKube Cloud (for SaaS-hosted tools), or a site-to-site tunnel for private endpoints.
  • About thirty minutes for a simple integration, longer for MCP.

​
Path A: Custom REST Integration

Use this when the tool exposes a REST API and an OpenAPI spec.
1

Gather the OpenAPI spec

Most internal tools already have one, or you can generate it from the service. RubixKube validates against OpenAPI 3.0 or 3.1.
2

Pick an auth type

Simplest option. Pass in a header or query parameter. Good for read-only or low-risk endpoints.
3

Register the integration

Open Integrations → Custom REST, upload the OpenAPI spec, select the auth type, and scope which environments can use it.
4

Test a single endpoint

The console has a test pane. Call one endpoint, confirm the response shape, fix any auth issues before enabling broadly.
5

Expose to the agent

Once green, the endpoints become tools the SRI Agent can call. Scope them in skill allowed_tools lists to keep the safety boundary tight.
Enable Require approval for any endpoint that creates, updates, or deletes data. Read-only endpoints can run unattended.

​
Path B: Custom MCP Server

Use this when the integration is richer than REST (think: a runbook framework, a bespoke deployment tool, or an internal data platform).
1

Write or adopt an MCP server

MCP is an open protocol. SDKs exist for TypeScript, Python, Go, and Rust. The server exposes a set of tools, each with a JSON schema for its arguments.
2

Host the server

Point RubixKube at a container registry and a manifest. We run the server for you in an isolated sandbox. Easiest for most teams.
3

Register in the console

Integrations → Custom MCP, then point at either the hosted URL or the manifest. Upload any required credentials to the workspace secret vault.
4

Validate tool schemas

The console lists every tool the server advertises. Click each to see its JSON schema and test it. Fix schema issues here, not in production prompts.
5

Scope to skills and agents

A custom MCP server’s tools are available to the SRI Agent once enabled. Scope them to specific skills via allowed_tools to avoid scope creep.

​
Auth types supported

Static credentials stored in the workspace secret vault. Rotated on a schedule you control.
Full handshake and refresh managed by RubixKube. Use this whenever the upstream tool supports it.
Client certificates for private APIs. Certificates are stored in the secret vault and attached per outbound call.
Webhook-style signed requests, JWT-in-header, and similar patterns. Configurable per endpoint.

​
What a good custom integration looks like

Narrow scope

One job, one tool schema per endpoint. The agent plans better against small, clear surfaces than sprawling ones.

Read-first

Ship read-only endpoints first. Add mutating ones once the read side proves reliable.

Approval on writes

Anything that mutates state needs human approval by default. You can relax later, never start permissive.

Typed arguments

JSON schema with enums and ranges beats free-form strings. Planning accuracy jumps.

​
Common questions

Yes. Point RubixKube at the internal URL via a tunnel, or ship a small runner in the same cluster as the Observer. No inbound firewall changes required.
Workspace secret vault. Encrypted at rest, scoped per integration, rotated on a schedule or on demand. Secrets never appear in agent prompts or logs.
No. Tool calls are not billed. Investigations count only whole detect-to-resolution cycles.
Enterprise customers can publish a custom integration as an internal template shared across their organisation’s workspaces. On lower tiers, configure each workspace separately.

How to add custom agent skills

Skills use these custom integrations as their tool surface.

Integrations overview

The out-of-the-box integrations and how they fit together.