MindsDB now connects agent work and model access through MindsHub

MindsDB’s current offer is MindsHub: an agent workspace for doing work and an inference layer for calling models through familiar APIs. That is a consequential shift for readers who remember MindsDB primarily as a database query engine. The company says the product was renamed in 2026 while the company and team stayed the same; the original query engine remains a community-maintained project.

01 / ProductThe current product is broader than the original database query engine

The official About page states the identity directly: MindsHub is the product and MindsDB is the company behind it. Its product-evolution explanation separates current MindsHub services from the original query engine. Coverage therefore uses one MindsDB company identity, with mindshub.ai as the current website, rather than treating a product rename as a second vendor.

The current site presents an agent workspace for research, analysis, reports and software work, alongside an inference product for developers. The workspace examples are demonstrations of the intended experience, not verified customer outcomes or tasks completed for this review. They show the kind of work the company wants the product to support.

The inference documentation describes Chat Completions, Responses and Messages as supported request formats. A developer can adapt an existing client to the endpoint instead of adopting a unique MindsDB SDK. That compatibility is an integration starting point; differences between the underlying models and API capabilities still matter.

02 / AudienceA candidate for teams that want model choice across practical work

MindsHub is worth evaluating when a team has several AI tasks and wants a deliberate way to choose models without making every provider change a full application rewrite. One task might summarize approved documents while another extracts structured fields or helps inspect code. The benefit depends on whether the shared access and workspace reduce work the team actually does.

The OpenRouter blueprint offers a relevant comparison for model access and routing. The Kong blueprint adds a gateway and operational-control perspective. Compare model catalogs, parameter behavior, observability and billing using the intended request patterns. A similar endpoint shape does not establish identical data handling or operational guarantees.

Teams specifically seeking the earlier MindsDB SQL query engine should treat it as a different deployment and support decision. The current About page says the original engine is maintained by the community. Do not assume that a hosted MindsHub account includes the old product’s connectors, commercial support or operating model simply because the company name is familiar.

03 / WorkflowProposed workflow: compare models on a bounded research brief

This proposed evaluation asks an agent to turn a small set of approved public documents into a structured research brief. It keeps the evidence set fixed so the team can compare model behavior and task cost without allowing a more expensive route to quietly collect a different body of sources. No confidential connector access is needed for the first pass.

1. 01

   Define the expected artifact

   Specify required fields, source citations, unresolved questions and a maximum scope. Keep a small human-written reference so omissions and unsupported statements can be identified consistently.

2. 02

   Select an explicit model route

   Choose a model or fixed version from the current catalog and record the requested identifier. Avoid a moving family alias during the initial comparison if reproducibility is the priority.

3. 03

   Run through a familiar request format

   Use the documented API shape already supported by the application. Check streaming, tools and structured-output behavior on the selected model before combining them into a larger workflow.

4. 04

   Inspect the returned evidence

   Validate the output structure in application code and check each material statement against the supplied documents. Preserve refusals or missing evidence instead of forcing every field into a confident answer.

5. 05

   Compare complete task consumption

   Record model tokens, tools, retries and elapsed time. Repeat the same task with another route and compare the quality of the resulting artifact as well as its invoice units.

The core-concepts guide explains that MindsHub translates requests to the configured provider; it does not host the underlying models. An unsupported parameter may be dropped and an out-of-range value clamped. The documentation names response headers for these adjustments, with limits on headers for streams whose response has already started. Inspect that behavior in the pilot rather than assuming every requested control reached the model.

The same guide distinguishes a moving alias from the model actually served and recommends logging both. This is useful for investigating a changed result after an alias update. A model switch should trigger the team’s own acceptance checks, even when the request format stays stable and the code does not require a new deployment.

For the workspace route, keep the same acceptance criteria: a finished-looking brief should expose its evidence and uncertainty. If the task later gains the ability to edit a repository or contact another system, evaluate those permissions separately. A successful research artifact does not establish that unrestricted external actions are appropriate for the next task.

04 / PricingThe hosted service uses a rate card and prepaid usage

The pricing page says Agents and Inference share the rate card and prepaid balance, without a subscription fee or minimum spend. The billing guide separately names a minimum wallet top-up. Those statements refer to different things: starting the service does not require a monthly commitment, while choosing to fund paid usage has a transaction minimum.

Do not estimate task cost from input tokens alone. The billing documentation includes generated reasoning in output-token accounting and distinguishes cache reads from cache writes. Search and page-fetch charges can also apply. A task that makes several model calls and uses external tools can consume more than the visible length of its final answer suggests.

The public free-tier language and detailed billing documentation describe allowance behavior at different levels. The fair-use policy excludes extensive automated machine use such as bulk pipelines and resale. Plan an automated production workflow around an approved paid or provider-key route, rather than treating the interactive free offer as unlimited backend capacity.

05 / DistinctionsAPI familiarity and a shared workspace address different kinds of friction

The inference layer is useful when the existing application already speaks a common model API. That can make a comparison easier to set up while keeping the caller’s basic interface familiar. It does not eliminate semantic differences: model context limits, tool behavior and image support still follow the selected route.

The agent workspace addresses a different problem: grouping work, inputs and resulting artifacts so a person can review a completed task. The current site presents both paths together. A team should evaluate whether it needs the workspace, the API, or both, because success in one does not establish the behavior of the other.

The explicit separation from the old query engine is also helpful. It prevents a buyer from treating historical database integration examples as a description of the current hosted product. Existing query-engine users can assess their deployment on its own terms, while new users inspect the current MindsHub documentation and support route.

06 / QuestionsProvider behavior and funding limits need a deliberate test

Confirm where the selected request is sent, what data the provider receives and how the application records that choice. A unified endpoint can simplify code while adding another service boundary. Keep secrets in the caller’s credential system and inspect the actual result metadata rather than assuming a stable alias identifies one fixed backend forever.

Exercise failed requests, exhausted allowance and an empty wallet in a controlled test. The billing guide documents distinct error paths and retry behavior. The application should explain when funding or rate limits prevent completion rather than repeatedly launching the same expensive task or presenting a partial artifact as finished.

For structured extraction, schema compliance is only the first check. A response can satisfy the requested JSON shape while giving an unsupported answer. Keep source-level validation and a human review path for consequential content. Switching models or enabling automatic fallback should preserve those same checks instead of treating availability as the only success criterion.

07 / DecisionChoose according to the current product you actually need