Integrations

No integration is hard-coded. Every external service follows the same 3-part pattern: Secret + Knowledge Doc + Primitive. Infinitely extensible without writing code.

The 3-Part Integration Pattern

Integration = Secret + Knowledge Doc + Primitive

1. SECRET          → /company/secrets/{service}.yaml
                     API key + allowed agents list

2. KNOWLEDGE DOC   → /company/knowledge/integrations/{service}.md
                     API docs, endpoints, conventions, examples

3. TOOL            → Agent calls http() or MCP server tool
                     Orchestrator auto-injects credentials

Why This Works

• Infinitely extensible -- add any integration without code changes
• Agent-native -- agents learn from knowledge docs, not hard-coded adapters
• Secure -- secrets are scoped per agent
• Debuggable -- every API call is logged
• No vendor lock-in -- switch from Linear to Jira by changing the knowledge doc

Integration Examples

GitHub (MCP Server)

GitHub is best integrated via its MCP server, which provides typed tools like create_pull_request, search_code, and create_issue.

Linear (fetch + GraphQL)

Linear uses a GraphQL API. Agents learn the query format from the knowledge doc and call it via fetch.

Slack (Webhook)

Slack can be used two ways: outbound messages via the Bot API, and inbound events via webhooks that trigger agents.

Stripe (http + Secret)

Stripe uses a REST API with form-urlencoded bodies. The secret provides the API key; the knowledge doc teaches agents the endpoint patterns.

When to Use MCP vs http

Integration Checklist

1. Add the secret -- autopilot secrets add <service>
2. Write the knowledge doc -- create knowledge/integrations/<service>.md
3. Set allowed agents -- decide which agents need access
4. (Optional) Configure MCP -- if an MCP server exists
5. (Optional) Add webhooks -- if the service sends events
6. Test it -- give an intent that uses the integration