Most search integrations only let an agent read. This one covers creating the index, loading documents into it, tuning relevancy and issuing API keys, so the whole loop happens in one session. The key you hand it decides how much of that loop the agent can actually reach — give it a scoped key rather than the master key when the job is only search.
Meilisearch's own server for its search engine. It connects to an instance you point it at and hands over the whole thing — indexes, documents, settings, API keys and the task queue — not just a search box.
- Search across a single index or several at once, with filtering, sorting and facets — search
- Indexes created, listed and deleted, with per-index metrics — create-index, list-indexes, delete-index, get-index-metrics
- Documents added or updated, and read back with pagination — add-documents, get-documents
- Index settings read and changed: ranking rules, faceting, searchable attributes, typo tolerance — get-settings, update-settings
- API keys created, listed and deleted with specific permissions — create-key, get-keys, delete-key
- Indexing work tracked and queued tasks cancelled — get-task, get-tasks, cancel-tasks
- Version, health, database statistics and system information — get-version, health-check, get-health-status, get-stats, get-system-info
- The server repointed at a different Meilisearch instance mid-session — get-connection-settings, update-connection-settings
- Every data-returning tool ships structuredContent alongside its text, so a client can read the result as data rather than parse prose
A running Meilisearch instance and Python 3.10 or newer; it starts through uvx, pip or the published Docker image. Two environment variables point it at the instance: MEILI_HTTP_ADDR for the URL and MEILI_MASTER_KEY for a key with the permissions you want the agent to have. Both can also be swapped mid-session — convenient in development, and the project's own advice is not to rely on that in production. MIT.
One command — uvx -n meilisearch-mcp