Connect your sources
Connect the sources you already operate. Tally reads their metadata, normalizes AI-relevant evidence, and preserves where each observation came from.
Add your first source#
- Open Connectors and select a category and connector.
- Give the connection a friendly name that identifies its account, region, or environment.
- Follow the form's Required Read-Only Permissions. Supply the requested endpoint, scope, and credentials, or workload identity where supported.
- Start with a bounded backfill window and record limit. Save and request a sync.
- Review the run result and follow an observation into Findings before expanding scope.
A requested sync makes work due for the background worker; it is not an immediate synchronous import. Leave an existing secret field blank to retain it when editing a connection. Credentials are encrypted before storage and are not returned by configuration reads.
Source categories#
| Category | Examples | Evidence |
|---|---|---|
| AI agents & platforms | ServiceNow, n8n, Copilot Studio, LangSmith/LangGraph, Dify, Bedrock/SageMaker, Azure AI, Vertex AI, and supported hosted-agent inventories | Saved agents, platform resources, or usage metadata, depending on the connector. |
| AI gateways & proxies | LiteLLM and Cloudflare AI Gateway | Provider/model requests and source-reported attribution. |
| Internet access & security | Defender, CrowdStrike, Zscaler, Palo Alto, NextDNS, Cato, Cloudflare, Rapid7, and Netskope | DNS, network, endpoint, and web activity supported by each source. |
| Infrastructure discovery | AWS/Azure/GCP network flows and resources, Kubernetes, Native Scanner | Network relationships and discovered runtimes/resources. |
The connector catalog and in-product setup form define the available types and fields. A vendor name in the catalog does not guarantee your account exposes every required API or event stream. Verify permissions and a real sync in your environment.
Understand what a source can prove#
Configured-agent inventory proves a saved definition, not execution. DNS or flow evidence shows contact with a service, not necessarily a model invocation. A client/framework label is a bounded source hint, not proof of a running SDK or coding agent.
LiteLLM reads its model-group and database-backed spend-log APIs. Enable spend logging and grant the required read access. Tally retains explicit user, requester IP, model, destination, and recognized client hints where available; it does not turn a proxy Host header into the requesting laptop's hostname.
Hosted-agent connectors retain bounded identity, display name, platform, state, and a literal configured model where supplied. They discard instructions, descriptions, execution content, and arbitrary tool configuration. An agent absent from a later poll is not immediately removed from historical inventory.
Rapid7 and Netskope setup#
Rapid7: use an Insight platform key with read-only Log Search access, the supported region, and the UUID of a normalized DNS Query log. The endpoint agent alone does not provide that log. The connector supports asynchronous queries and paginated results; it does not ingest raw Endpoint Agent process logs.
Netskope: use a commercial https://<tenant>.goskope.com origin and a v2 token with read permission for the selected page or application Events API. Create separate connections for both event types. Web transaction streaming and the stateful dataexport iterator are outside this connector's scope.
These connectors use bounded, settled time windows with a five-minute delay. Cursor progress is saved only after ingestion succeeds; retries preserve event identity. Source retention and late-arriving events can still limit coverage.
Uploads and authorized scanning#
Supported Defender, Zscaler NSS, and Palo Alto exports use a browser upload flow: choose an authorized JSON/NDJSON export, preview the canonical digest and counts, authorize that preview, then commit it. The raw uploaded file is not retained.
For continuous receiver integrations, inspect the packaged tallyd collector receive --help and source-specific adapter contract. Browser uploads do not configure a streaming bridge automatically.
Native Scanner requires an explicitly authorized scope. Preview its canonical targets and ports before acknowledging authorization and starting. Results are resource-discovery evidence; an open service is not counted as an inference request.
Maintain a connection#
Use recent connector runs and Audit logs to diagnose permissions, source errors, and stalled cursors. Start with a bounded backfill when new attribution fields become available; previously discarded raw data cannot be recovered.
Deleting a connection removes its configuration, schedules, and sync receipts. It does not delete all historical findings. Use scoped data deletion when retained operational data must also be removed.