The people shaping Meilisearch 0.29 surge for teams choosing search engines

Meilisearch updates and the team behind the engine

Meilisearch version 1.54 introduces a native Model Context Protocol endpoint at /mcp. This allows AI agents to query indexes directly using an API key through bearer authentication. Quentin de Quelen, the CEO and co-founder, presented AI-powered search and discovery at IBC2026 in Amsterdam on September 11. He demonstrated how hybrid search combines full-text and semantic matching to serve results from a single index. The company employs 8 people and raised a €1.5m pre-seed round. Meilisearch also added experimental search rules to boost, demote, or hide results. A scale action with a weight above 1.0 boosts seasonal items, while a weight of 0.0 hides discontinued products. Version 1.51 improved search speed by 5.4x for datasets with more than 14,000 fields. Version 1.49 made synonym loading faster, which increased search speeds by up to 13x for indexes with large synonym lists. Version 1.53 allows foreign filters to work across shards, enabling retrieval of up to 1,000 documents from a related index in sharded setups. Cloud projects can upgrade to new versions automatically for XS, S, and Build plans.

Infrastructure decisions and memory management

Infrastructure decisions determine whether a team chooses Meilisearch, Typesense, or Algolia. Meilisearch provides a single Rust binary with a memory-mapped LMDB store on local disk, which relies on the kernel page cache to decide exactly what stays resident in the process address space. If the hot working set does not fit in the page cache, query latency becomes unpredictable. Typesense uses an in-memory approach where the index stays in RAM. This means RAM requirements act as a direct function of index size. You know the basics of search indexing, so evaluate the trade-off between memory and disk. To prevent an Out of Memory error, you should keep at least 15% free memory for the operating system.

Self-hosting Meilisearch on a virtual machine costs only the price of the machine. This option gives teams control over data residency and eliminates per-search pricing. However, Meilisearch is a prefix-based engine. It cannot find parts of a query that lie in the middle or at the end of a word. Security requires keeping the master key in the server environment rather than a browser. Admins should use an admin key for the backend indexer and search-only tokens for clients. A reliable recovery plan requires a script to rebuild the index from the source database.

The shift from commercial search to open source

Teams switch from Algolia to avoid usage-based pricing and closed-source licenses. Algolia scales costs based on search volume and record counts. Meilisearch and Typesense offer free alternatives for self-hosting. Meilisearch has 60,000 GitHub stars, while Typesense has 27,000 stars. Meilisearch focuses on typo-tolerant search-as-you-type. The ranking rules in Meilisearch use an ordered list of weights for words, typo, proximity, attribute, sort, and exactness. Version 1.54 also includes an optional inlineRule parameter for search query objects. This feature requires Meilisearch Enterprise Edition and the dynamicSearchRules experimental feature. One recent fix in version 1.54.3 prevents errors when disabling binary quantization in embedder settings.

One engineer noted that the migration to Elasticsearch from Algolia was driven by price and the closed-source license. Using a search engine you own allows you to avoid outsourcing a core business component. Meilisearch is optimized for scenarios where a user types into a box and expects results before finishing the word. It is not a log store or an analytics engine. It is a search index over data that lives in a separate database. Does the need for absolute control over ranking rules outweigh the operational effort of managing a single-node instance?