Query MongoDB

An inventory assistant uses Bash and Python's pymongo to find products that need reordering. The database connection string lives in the worker container's environment. The application sends a question to Mango and prints the agent's answer.

The example source uses Mango's Go SDK for Agents, Sessions, Events, and Environment Work. Its own Docker launch passes MONGO_URI into the container. This is application-owned configuration; it needs no new Mango API field or change to the reference worker.

Run with a local database

You need Go, Docker with Compose, and a running Mango deployment with a configured model. Use the same source revision for the deployment and SDK. These commands run from the repository root.

Build the example worker and start its separate MongoDB container:

docker build -f examples/mongodb-query/Dockerfile -t mango-mongodb-example:local .
docker compose -f examples/mongodb-query/compose.yaml up -d --wait

The database contains three synthetic inventory records and exposes no host port. The worker reaches it on the mango-mongodb-example Docker network.

Configure the application and run it:

export MANGO_BASE_URL=http://localhost:8080
export MANGO_DOCKER_BASE_URL=http://host.docker.internal:8080
export MANGO_API_KEY=your-mango-workspace-key
export MONGO_URI=mongodb://mongodb:27017
scripts/with-dev-env go run ./examples/mongodb-query

scripts/with-dev-env loads MANGO_MODEL_ID from your local development config. Alternatively, set that variable yourself and run go run ./examples/mongodb-query. The model provider's credentials belong to the Mango orchestration worker; the example only needs the model ID and Mango Workspace key.

MANGO_DOCKER_BASE_URL must be reachable from the container. For a local Mango API, bind it to an address Docker can reach, as described in the self-hosted worker guide. On Linux the example adds the standard host.docker.internal:host-gateway mapping.

The default question asks which products need reordering. You should see a Bash tool call and an answer with:

• TEA: stock 4, reorder point 12; order 8.
• MUG: stock 2, reorder point 6; order 4.
• COFFEE needs no order: stock 18 is above its reorder point of 10.

The wording can vary. Inspect the tool output and answer as functional evidence; the example is a small application, not a runtime conformance or recovery test.

Verified on 2026-09-13 with claude-sonnet-4-5-20250929, the local MongoDB above, and Mango's self-hosted worker SDK. The Bash output and final answer both identified MUG (4 units) and TEA (8 units). This verifies the inventory query journey; it does not establish Atlas search or other database workflows.

How the worker runs

The application creates a dedicated Environment and Session. The SDK's WorkPoller polls and acknowledges one Work item; the application launches a container and passes the scoped Work payload through stdin. Inside that container, EnvironmentWorker.HandleItem owns heartbeats, tool execution, result delivery, and Work Stop. The application keeps reading Events across requires_action pauses until the agent ends its turn.

MONGO_URI is passed with Docker's --env MONGO_URI. Its value is absent from the Agent definition and user message. The agent's shell can read it, as it must to connect; use a database account with the permissions appropriate for your task. The Workspace API key is not passed to the container.

Use your own inventory collection

Set MONGO_URI to a MongoDB connection string reachable from the container. The image includes SRV support for mongodb+srv:// connections. Your collection should have sku, name, stock, and reorder_point fields. The example queries it and does not seed or remove records from your database.

Cleanup

The application removes its worker container, deletes its Session and Environment, and archives its Agent on exit. Its workspace is temporary. It leaves the database running so you can inspect or query it again.

Remove the example database and its disposable data when finished:

docker compose -f examples/mongodb-query/compose.yaml down --volumes

If the host application is forcibly killed, its printed Session ID and Docker container name (mango-mongodb-<work-id>) identify resources to clean up manually.

The scenario follows the direct-database-access pattern in CMA's Docker self-hosted example. Mango runs its own control plane and worker SDK. This example does not implement the separate Atlas fraud-review, search, or human-approval cookbook.