Open source and self-hosting options for agent context

Security teams often ask the first question in a catalog evaluation: can we run it ourselves? For agent context the question matters more, because the context layer sees schema, lineage, usage and sometimes sample values for every critical system. Here is what the vendors we reviewed document on their public pages.

DataHub offers DataHub Core as open source and DataHub Cloud as a managed enterprise product. DataHub says the open-source project is used by more than 3,000 organizations and has a community of over 13,000 members, with deployment options from a single node to cloud-hosted, hybrid and decentralized setups.

Source: DataHub Series B announcement

Cube's homepage shows 21,018 GitHub stars and 12,830 Slack members for its open-source semantic layer, with Cube Cloud as the commercial platform.

Source: Cube homepage

Atlan publishes an open-source App SDK under Apache 2.0 for building custom connectors, though the platform itself is offered as a service.

Source: Atlan connectors

Stardog lists Stardog Enterprise for on-prem alongside Stardog Cloud and cloud marketplace offerings, and a free edition that is not open source. We did not score Stardog because our sources did not cover enough criteria.

Source: Stardog pricing

Honeydew processes data inside your data warehouse using your own LLM provider. That is not self-hosting the product, but it keeps the data processing in your environment. AtScale computes queries inside your data platform.

Source: Honeydew homepage

Source: AtScale: semantic context for AI via MCP

Euno describes itself as a cloud-native SaaS platform that reads metadata only. Atlan states it never queries the rows in your tables and does not state a self-hosted option on the pages we read. Alation, Collibra and Select Star do not state deployment options on the pages we read.

Source: Euno homepage

Source: Atlan connectors

Self-hosting moves operational work to your team: upgrades, scaling, connector maintenance and security patching. It gives you control over where metadata lives and how it is accessed. For agent context, also consider freshness: the value of the layer depends on how quickly it reflects changes, and a self-run deployment is only as fresh as the jobs you schedule.

A metadata-only SaaS can be a reasonable middle ground. If the vendor reads schema, lineage and query logs but never rows, the sensitive surface is smaller. Ask each vendor for the exact list of metadata it stores, where, and for how long, and have security review it before the proof of concept.

Even with SaaS, you can narrow what leaves your environment. Ask whether the vendor can connect through a private network link, whether it needs write access anywhere, and whether query-log ingestion can exclude sensitive schemas. Ask for the list of sub-processors and the regions where metadata is stored. DataHub lists SOC 2 Type II, a 99.5% uptime SLA, RBAC and SSO on its context platform page, and Select Star lists SOC 2 covering security, confidentiality and availability; for every vendor, ask for the current report rather than relying on a badge. Security reviews for this category often take longer than the proof of concept itself, so start them in parallel. A metadata-only tool still sees table and column names, query text and usage patterns, which can reveal a lot about a business, so treat the review as seriously as for any system holding customer data.

Source: DataHub Cloud context platform

Source: Select Star homepage

Deployment options carry 8% of the weight, and each review shows the deployment score with its reason. If self-hosting is a hard requirement for you, use the open source or self-hosting filter in the weight-your-own tool rather than raising the weight: a requirement should remove options, not just reorder them.

Finally, revisit the question each year. Vendors add deployment options as enterprise demand grows, and a SaaS-only product today may offer a private deployment later. Record the answer you got, the date and the person who gave it, so the next review starts from facts rather than memory.