2026-11-14 · ← News
Datasette 0.1a11 connects local models through the OpenAI interface without another plugin
datasette-llm 0.1a11 can connect directly to local and hosted models through an OpenAI-compatible interface. Operators gain more choice, along with responsibility for endpoint addresses, credentials and the actual enforcement of model allowlists.
The image could not be loaded.
datasette-llm 0.1a11 adds direct support for OpenAI-compatible endpoints. Datasette can now use local models in LM Studio, Ollama or llama.cpp as well as remote services without requiring a separate provider plugin for each one.
One configuration object now describes local and remote models
A model can be defined as an object containing a local name, the model name expected by the server and an api_base address. The release supports OpenAI Chat Completions and Responses, tools and streaming. A new endpoints setting registers these models separately without enabling the models allowlist.
Credentials are specified with key. Datasette looks in datasette-secrets first and then in stored LLM keys. An endpoint without a key is treated as unauthenticated and does not receive another provider's default credential. The maintainers explicitly warn against putting secrets in headers because extra headers appear in the configuration output. The release requires LLM 0.37.1 or newer.
OpenAI compatibility turns the provider into a configuration choice
For teams using Datasette, the cost of trying a local model falls. The same plugin can point to a laptop running Ollama, an internal server or a hosted service as long as they speak the same protocol. The practical gain is that other Datasette plugins no longer need to know every provider individually.
The change also separates a model's local identity from the name expected by the remote server. That makes backend swaps easier without rewriting the entire interface, but only where protocol implementations genuinely match.
A shared protocol does not guarantee identical backend behavior
OpenAI-compatible describes an interface, not model quality, complete feature parity or operational security. Tools, streaming and Responses can differ at the edges between servers. A local endpoint protects data only when the configuration really points to a local service and its logs remain under the operator's control.
The release also tightens filters. Allowlists, blocklists and the llm_filter_models hook now apply to direct model selection, while a missing required key is reported during selection. This follows the issue fixed in 0.1a10, where the model list in the demo did not match its configuration.
Operational tests must exercise every selected endpoint
The next useful signal will be tests that run the same tasks against different backends. A successful response is insufficient. Teams need to check tool calls, streaming, error paths, limits and whether a blocked model can be reached through another route.
Version 0.1a11 remains an alpha. Routine use will decide whether one configuration format really reduces integration work or merely moves provider differences into runtime debugging.
Lilith's verdict
Datasette has put Ollama, LM Studio and hosted services on one platform. It now has to ensure every train follows the same timetable and none quietly departs with someone else's credential.