Connect tools and capabilities to Managed Deep Agents

Connectors extend an agent with external tools and capabilities, remote MCP tools, constrained LangSmith operations, and GitHub sandbox access, without wiring up your own clients, OAuth flows, or credential plumbing. Managed Deep Agents discovers connector modules under connectors/. Each file directly under that folder is a connector; you do not register connectors in the agent entry (agent.py or agent.ts). Every connector module exports a named connector created from the connectors namespace (for example connectors.mcp(...)); a TypeScript export default fails the build.

​

Connector types

For the full project layout, see the CLI project file reference. To scaffold a connector and complete its workspace OAuth from the terminal, use mda connect (for example mda connect gmail). Run mda connect --list to see every connector and its connection status.

​

Choose the right integration

For example, the GitHub connector prepares repositories in a sandbox and can also receive App webhooks through its events option.

​

Provider events and messaging

Some connectors also receive inbound provider events and reply on the same conversation. The Slack connector handles Slack Events, and the GitHub connector handles GitHub App webhooks. Both mount a public ingress route on the Agent Server at POST /connectors/{name}/events, verify the provider signature, invoke the agent with identity stamps (source.provider), and can auto-reply on the originating conversation. Provider events require a root identity declaration. For a shared workspace bot, scope threads by the provider source conversation with define_identity(scope={"threads": "conversation"}). To share history between a browser and Slack, keep threads user-owned and link accounts with Connect-with-Slack. Provider-specific delivery, secrets, and setup live on each connector page.

​

Combine connectors with authored tools

Use custom tools for business logic, private APIs, database access, and other project-owned code. Use custom middleware for cross-cutting behavior around model calls, tool calls, lifecycle hooks, retries, limits, and data handling. MCP connector tools are appended to the tools you define in the agent entry. LangSmith capabilities are exposed on separate HTTP routes scoped by identity.

​

Test and deploy

Test the project locally with mda dev, then deploy it with mda deploy. Open deployment traces in LangSmith to inspect model calls, tool calls, errors, and latency. Connector misconfiguration surfaces during local startup or first tool load. LangSmith capability calls return 401 without a resolved identity and 403 when ownership checks fail. For deploy symptoms and fixes, see Troubleshooting.

​

Next steps

---