Build your Data Store

Published on August 6th, 2026

Introduction

This page covers the path from an empty Context Store to working Capabilities. The app tracks the same steps on its Getting started page, and marks each one complete as you go. You can do them in any order.

The six steps are:

  1. Connect an Integration
  2. Add Data Sources
  3. Ingest data
  4. Link objects with Relationships
  5. Create a Context Group
  6. Create a Capability

1. Connect an Integration

An Integration is a connection to an external system. It holds the configuration and credentials that its Data Sources use, and it enforces rate limiting across all of them.

Go to Integrations and pick a pre-built Integration, or use + New to define your own.

Most pre-built Integrations need only a secret before they work. Secrets are managed under Administration → Secrets, and you can create one inline while configuring the Integration.

Check what data you intend to pull before you mint a credential. Many third-party tokens gate access per resource type, so a token that can list repositories may not be able to read their dependency alerts. Pre-built Integrations offer an Automatically enable data sources toggle. If you leave it on, each of the Integration's pre-built Data Sources is enabled after saving, once it passes a dry-run. That covers step 2 for you.

2. Add Data Sources

A Data Source is a scheduled job that calls an endpoint, or a chain of endpoints, and stores what it returns as objects in the Data Store.

Go to Data Sources. Enabling an Integration in step 1 will normally have populated this list already. Enable the ones you want, or use + New to build a custom Data Source.

A Data Source can enrich its results with a second call. For example, fetch a list of GitHub logins, then loop over that list to fetch the full user entity for each one. See the Data Store overview for how chaining works.

3. Ingest data

Enabling a Data Source schedules it. To see results immediately, run one from the Data Sources page.

Once a run completes, its objects appear under Explore → Data Store, where you can confirm you pulled what you expected. See Exploring the Data Store.

If a run returns nothing, the usual causes are a credential without the right permission, or a filter on the Data Source that is narrower than you intended.

Relationships are the edges of the graph. They are what lets an agent walk from a repository to the team that owns it, and from there to the person on call for it.

Go to Relationships. Enabling a Data Source seeds suggested rules for you. You can also:

  • Draw connections manually between Data Sources
  • Accept algorithmic suggestions
  • Accept suggestions generated by a connected LLM

See Relationships for the detail, and Graph view to check the result visually.

5. Create a Context Group

A Context Group collapses several objects into one concept. An Employee group might combine a GitHub user, an AWS IAM identity, and a BambooHR record, so that an agent looking up a person gets one record rather than three partial ones.

Go to Context Groups and define a rule. Like Relationships, these are seeded automatically when you enable Data Sources. See Context Groups.

6. Create a Capability

A Capability is a documented procedure an agent can follow, written as structured markdown. Runbooks and skills both fit the shape.

Go to Capabilities and create one. Capability instructions can reference the things you built above inline, using @type:slug:

  • @datasource:sentry-projects
  • @context-group:payments-team
  • @action:create-repository
  • @capability:deploy-service

These references are also read by the Service Token scope picker, which uses them to suggest the scopes a Capability needs.

See Capabilities.

Connect an agent

With data in the graph, expose it over MCP.

  1. Enable the servers you want under Administration → MCP Servers. explore is on by default. integrations, manage, and actions are not.
  2. Point your agent at the endpoint. Use https://app-api.roadie.so/api/mcp/v1/ for the union server, or the URL of an individual server.
  3. Authenticate with OAuth, or with a Service Token for non-interactive clients.

Tool calls made against your Context Store are recorded under Agent Sessions.

Next steps