OpenSearch MCP Server
for AI Agents

Index a document at a specific ID; overwrites if the ID already exists unless op_type=create.

Attach a single alias to a single index (non-atomic — use update_aliases for atomic swaps).

Compact tabular listing of every index — health, status, docs count, disk usage, shards.

Compact tabular alias listing — alias, index, filter, routing, is_write_index.

Snapshot of cluster health — status (green/yellow/red), active/relocating shards, and pending tasks. The canonical connection-health check.

Aggregated statistics across the whole cluster — nodes, shards, storage, JVM heap, and OS/process metrics.

Full cluster state — routing table, metadata, node roster, and index settings. Very large payload; filter via the `metric` + `index` path segments.

Tasks queued on the cluster manager but not yet executed — useful to spot backlogs during allocation storms.

Fetch a single document (source + metadata) by its ID.

Fetch only the _source field of a document, without _seq_no / _primary_term / _version metadata.

Return the settings, mappings, and aliases of one or more indices.

Return details of a specific alias on a specific index (or across a wildcard set).

Return every alias across every index (or a filtered subset).

Same as search but without an index in the URL — required when using a Point-in-Time reference.

Return every registered snapshot repository on the cluster.

List snapshots in the given repository — companion LIST for the create/get/delete snapshot actions.

Return details of one or more snapshots — status, indices, start/end times, shard counts.

Fetch the status of a running or completed task by task_id.

Delete a document by ID.

Remove an alias from an index.

Delete a snapshot and reclaim its unique segments from the repository.

One-line summary of cluster health — status, node counts, shard counts.

Per-node overview — heap, CPU, load, disk, role, and cluster-manager status.

Search and paginate Dashboards saved objects. The canonical enumeration endpoint for datasync discovery.

Fetch many saved objects in a single request.

Add a document to an index and let OpenSearch generate the ID.

Existence check for a document — returns 200 if the document exists, 404 if not.

Apply a partial update to a document using the update DSL (doc merge, script, or upsert).

Execute many index/update/delete operations in a single request. **Body must be NDJSON** (newline-delimited JSON), not standard JSON — each operation is metadata line + optional source line.

Same NDJSON bulk semantics as bulk_operations, but scoped to a single index in the URL — operations can omit `_index` in their metadata.

Fetch multiple documents by ID in one round-trip, optionally across different indices.

Copy documents from a source index (or remote cluster) to a destination index; the canonical datasync migration primitive.

Existence check for an index — 200 if it exists, 404 if not.

Force a refresh so recently indexed documents become searchable immediately.

Persist in-memory transaction log to disk.

Merge index segments to reduce their count; heavy operation.

Re-open a closed index so it becomes readable/writable.

Close an index so it stops accepting reads/writes; frees cluster resources.

Apply multiple alias add/remove operations atomically.

Run several searches in one round-trip; body is NDJSON with alternating header + query lines.

Return the count of documents matching a query, without returning the hits themselves.

Fetch the next batch of a scroll context. Discouraged for new work — use PIT + search_after instead.

Release the memory held by one or more scroll contexts.

Register a snapshot repository backed by fs, s3, gcs, or azure storage.

Create a snapshot of one or more indices to the target repository.

Restore indices from a snapshot back into the cluster.

Cancel a running task by task_id. Cancellation is cooperative — the task ends at the next check point.