Two open-source data agents make refusal part of the product contract

Two open-source data tools shipped releases this weekend from opposite ends of the analytics stack. Jeeves v1.0.0 puts a natural-language query widget inside Laravel applications. Dex v1.11.0 adds native Apache Ossie semantic-layer support to an agent-oriented analytics-engineering toolkit. The shared signal is not simply broader integration. Both projects document where their automation stops.

Jeeves treats the database principal as the final control

Jeeves sends table names, column names, types and user-authored semantic descriptions to a model, then executes the returned SQL locally. Its release notes say result rows are not sent upstream, and the repository includes an end-to-end privacy test using sentinel values to check that claim across query paths.

The more important caveat is in the security documentation: the SQL validator is pattern-based, not a full multi-dialect parser. Jeeves therefore requires generated SQL to use a separately configured database connection rather than the application's default connection. The project tells operators to grant that principal SELECT only; its diagnostic command attempts a write on PostgreSQL and MySQL to verify that the database rejects it.

That is a stronger contract than presenting validation as a sandbox. A pattern filter can miss a construction its authors did not anticipate. A genuinely read-only principal makes that miss a failed statement rather than a write.

Jeeves is also unusually direct about answer quality. The project reports 35 correct answers out of 46 on its own 14-table set and 30–31 out of 36 on a Spider development sample, each across three Gemini 2.5 Flash runs. Those small, project-run measurements are not an independent benchmark. They are still useful because the README translates them into an operational warning: roughly one question in four was wrong on the uncurated internal schema, with three-table questions scoring 2/6 in each run.

Dex preserves the difference between interchange and execution

Dex's new Apache Ossie adapter reads and authors native Ossie documents, incorporates declared keys and relationships into warehouse exploration, and tracks definition drift. Ossie itself describes a vendor-neutral JSON/YAML specification for exchanging semantic metadata across analytics, AI and BI systems.

Dex explicitly refuses to execute a metric by name through an Ossie layer. Its compatibility statement says Ossie does not define a portable query runtime, filter grammar, join planner or execution semantics. Rather than inventing those pieces, Dex exposes catalog and authoring operations while declining the unsupported query.

The implementation also pins the upstream Ossie schema by content hash. Its published compatibility matrix distinguishes hard validation errors, warnings and skipped checks, and names known differences from the moving upstream schema. That does not prove interoperability with every Ossie consumer; Dex says so.

Refusal is evidence, not missing polish

Jeeves and Dex are young projects with modest adoption signals: at verification, Jeeves had no GitHub stars or forks, while Dex had 24 stars and nine forks. Neither release should be read as market validation.

But both offer a useful evaluation test for data agents: ask not only what the tool can generate, but what authority remains outside the model. In Jeeves, the database principal owns the irreversible boundary. In Dex, the source specification owns the semantic boundary. A tool that can name and enforce its refusal path is easier to assess than one whose limits appear only after a wrong query runs.