Context Data

Organizations running AI agents against their own databases, files, and APIs face problems that tool-permission systems don't solve: agents re-fetch the same records constantly, can't find data by content across large document sets, and can execute destructive writes (dropped tables, unbounded deletes, schema changes) that role permissions alone don't catch. Existing identity and tool-governance platforms decide which tools an agent can call, but nothing sits at the data layer itself to cache repeated access, redact sensitive fields before they reach an agent, or hold a risky write for human review. Teams also lack visibility into why an evaluation metric moved after a dataset or pipeline change, and lack a full trace of what an agent actually did once a production run goes wrong.

Onyx is a runtime that sits between agents and an organization's databases, files, and APIs, speaking the protocols agents already use (Model Context Protocol, Postgres wire protocol, HTTP) so adoption is a connection-string change rather than a code change. It resolves identity, applies role and classification-based policy, serves repeat queries from a deterministic and semantic cache, redacts sensitive fields in the result stream, and logs every request to an immutable audit trail; destructive writes are held for human approval instead of executing automatically. Alongside Onyx, Cleanroom attributes evaluation metric movements to specific dataset or pipeline changes and flags benchmark contamination, while Chronicle records full distributed traces of agent runs (steps, tool calls, decisions) for debugging and audit. All three products are open source and can be self-hosted or run as a managed service.

As agents become how organizations use their own data, the page argues that data access needs dedicated infrastructure the same way tool access already has identity and governance platforms.