MindsDB

MindsDB has evolved from an AI layer for databases into a broader connected-data agent platform. The current homepage presents “Cowork Cloud” agents that can research, analyze, create reports, build apps, and schedule recurring work. Its strongest examples are concrete: combine Gong calls, Zendesk tickets, and HubSpot notes, cluster customer-feedback themes, then rank them by linked revenue rather than raw frequency; or prepare a morning brief that links every action item back to its Slack thread or email. Connectors shown on the site include PostgreSQL, HubSpot, Gong, Zendesk, and Notion.

The platform’s security design is relevant to agent builders. The vendor says connection secrets are stored in a vault and scoped per connection, so an agent does not receive a raw key. Project memories can preserve instructions such as using read-only database credentials, treating currency as USD, or leading reports with metrics. Source links in generated reports help a reviewer trace a conclusion back to the underlying ticket, call, or record. MindsDB can also expose connected capabilities through an MCP server, although teams should verify authentication, transport, and write permissions for the exact version deployed.

Pricing and evaluation

The live pricing page offers the first 5 million MindsHub Air tokens each month at no charge in both Cowork and the API. Beyond that allowance, the page listed $0.20 per million input tokens, $1.20 per million output tokens, $0.02 per million cached-input tokens, and $0.25 per million cache-write tokens. Long prompts over 272,000 tokens were listed at $0.40 input, $1.80 output, $0.04 cached input, and $0.50 cache write per million tokens. The vendor says Air is an alias that tracks its preferred balance of cost, speed, and intelligence, so the target model and rates can move. Third-party models have their own rates; for example, the fetched page showed Anthropic web search at $10 per 1,000 searches. These details warrant manual verification before budgeting.

A pilot should connect read-only copies of one CRM, support system, and database. Give the agent five questions with known answers and five ambiguous requests. Score citation accuracy, join correctness, freshness, p95 completion time, token use, and whether it obeys row- or source-level permissions. Then schedule one report and verify duplicate prevention, failure alerts, and the effect of a revoked credential. Do not grant production write access until approval and rollback behavior are proven.

MindsDB differs from a pure warehouse or chat interface by combining connectors, persistent instructions, scheduled agent work, and source-linked outputs. Compare governed lakehouse agents in Databricks Mosaic AI, document retrieval through GroundX, search infrastructure from Jina AI, and request-cost observability in Helicone. The vector database guide is useful if retrieval architecture is the real decision.

The key drawbacks are operational. Costs cross token, cache, model, and search dimensions. An alias that changes models may complicate reproducible evaluations. Cross-system agents can also overreach if scopes are broad. MindsDB is valuable when linked business data and recurring analytical work are the bottleneck, provided the team treats permissions, citations, and spend limits as core implementation work.