Agent Discovery Integrations

Arthur finds the AI agents running across your organization by connecting to systems you already use. This page is the configuration reference for each integration: what it discovers, the fields it takes, and the permissions its credentials need.

How a discovery source is configured

A discovery source has two parts:

• Source fields: how Arthur connects to the system: its URL, project or account, and credentials. Fields marked Sensitive are stored as secrets and never appear in logs or error messages.
• Scan configs: what each scan does. A source can have more than one scan config. Each one has:

Every integration accepts only https:// URLs.

Available integrations

Google Cloud Vertex AI Agent Engine (gcp_vertex)

Lists every agent deployed to Vertex AI Agent Engine in one GCP project and region. To cover more regions or projects, add one source for each.

Source fields

* Required unless your Arthur engine runs on GCP and has been configured to use its own Application Default Credentials (ARTHUR_ENGINE_GCP_VERTEX_ALLOW_ADC=true).

Scan config

Required permissions

The service account needs only aiplatform.reasoningEngines.list on the project. The predefined role roles/aiplatform.viewer includes it. For least privilege, grant a custom role with just that permission.

Google Cloud Compute Engine (gcp_compute_engine)

Discovers agentic coding tools, local agents, AI browsers and browser extensions installed on your Compute Engine VMs. The Arthur discovery collector runs on each VM and writes its results to the VM's guest attributes, and Arthur reads them through the Compute Engine API.

Before you start

1. Deploy the Arthur discovery collector to your VMs. It runs on a schedule through an OS policy.
2. Enable guest attributes by setting the metadata key enable-guest-attributes=TRUE on each VM or on its project. Arthur reports VMs without it as guest-attributes-disabled and doesn't read them.

Source fields

* Required unless your Arthur engine runs on GCP and has been configured to use its own Application Default Credentials (ARTHUR_ENGINE_GCP_COMPUTE_ENGINE_ALLOW_ADC=true).

Label matching is exact. If no VM has a label in your selectors, the scan logs a warning and continues; it doesn't fail.

Scan config

Required permissions

• compute.instances.list
• compute.instances.getGuestAttributes
• compute.projects.get. Needed only when guest attributes are enabled at the project level rather than on each VM.

The predefined role roles/compute.viewer includes all three.

Jamf Pro (jamf_pro)

Discovers agentic coding tools, autonomous local agents, AI browsers and browser extensions on your managed Macs. The Arthur discovery collector runs on each Mac and reports through a Jamf Extension Attribute. Arthur reads those results from the Jamf Pro API. Arthur makes read-only API calls and never runs anything on a device.

Before you start

1. Create the Arthur Extension Attribute in Jamf Pro.
2. Deploy the Arthur discovery collector with a Jamf policy.
3. Create an API role and API client in Jamf Pro (Settings → API roles and clients) with the privileges below, and generate a client secret.

Source fields

A group name that doesn't exist in Jamf Pro fails the scan rather than being skipped. Otherwise a mistyped group name could quietly widen or empty the scan.

Scan config

Required permissions

The API role needs:

• Read Computers
• Read Smart Computer Groups and Read Static Computer Groups. Needed only if you set include_groups or exclude_groups.

Splunk Enterprise (splunk_enterprise)

Runs your SPL query as a Splunk search job over the lookback window. Each row the query returns becomes one discovered agent. Your query decides what counts as an agent: gateway logs, proxy logs, EDR events, or anything else Splunk indexes.

Source fields

Scan config

Required permissions

The token's role must be able to run searches and must be allowed to search every index your query reads (srchIndexesAllowed). If a scan returns zero rows, check index access first.

Elastic Security (elastic_security)

Runs your ES|QL query against Elasticsearch over the lookback window. Each row the query returns becomes one discovered agent. Query DSL isn't supported.

Source fields

Scan config

Elasticsearch caps an ES|QL result at 1,000 rows by default, or 10,000 with LIMIT 10000. If a result reaches the cap, the scan logs a warning.

Required permissions

The API key needs only read on the indices your query names. No cluster privileges are needed.

Example: self-hosted Kong Gateway

Self-hosted Kong Gateway has no usage API, but if its http-log plugin ships to Elasticsearch and ai-proxy logs statistics (logging.log_statistics: true), this query lists one agent per Kong consumer and the models it used:

FROM kong-ai-logs
| WHERE consumer.id.keyword IS NOT NULL AND ai.proxy.meta.plugin_id.keyword IS NOT NULL
| EVAL model = CASE(
    response.status < 400 AND ai.proxy.meta.response_model.keyword IS NOT NULL,
    CONCAT(ai.proxy.meta.provider_name.keyword, "/", ai.proxy.meta.response_model.keyword),
    NULL)
| STATS last_seen = MAX(@timestamp),
        name = MAX(consumer.username.keyword),
        llm_models = VALUES(model)
    BY consumer_id = consumer.id.keyword
| EVAL external_id = CONCAT("kong:kong:8000:", consumer_id),
       name = COALESCE(name, consumer_id),
       llm_models = MV_SORT(llm_models)
| KEEP external_id, name, last_seen, llm_models
| SORT external_id
| LIMIT 10000

Replace kong-ai-logs with your index and kong:8000 with your Kong proxy's host. Once you've scanned, don't change the external_id prefix: it's each agent's permanent identifier.

Arthur Evaluation Engine OTEL sink

Every application or agent that sends OpenTelemetry GenAI traces to the Arthur Evaluation Engine is included in discovery automatically, including traces sent without a service name. There's nothing to configure beyond sending traces.

SIEM query output contract

Splunk and Elastic queries must return one row per agent, using only these column names. If a query returns a column not listed here, the scan fails as not_configured. Rename helper columns or drop them.

A row with a blank external_id or name, or a last_seen that isn't a timestamp, is skipped and logged. The rest of the rows are still published. Leave out optional columns your data can't supply rather than returning them empty.

Coming soon

Don't see a system you use? Contact us. We prioritize new integrations based on customer demand.