Database

Database extension

Database is a native workbench for relational, key-value, document, search, and analytical databases. It exposes the same validated operations through the Edith interface, the ed database command family, and a local MCP server.

The first supported product set is:

• PostgreSQL
• MySQL
• MariaDB
• SQLite
• Redis
• Valkey
• MongoDB
• Elasticsearch
• OpenSearch
• ClickHouse

Support is product-specific. A shared protocol does not imply that metadata, types, permissions, administration, or failure behavior are interchangeable. Every live connection produces a capability report based on the detected product, version, topology, modules or plugins, and current account permissions.

Product principles

1. A user always sees the active connection, environment, namespace, and mutation mode before execution.
2. The interface, CLI, and MCP server call one typed command executor. Validation, capability checks, safety checks, redaction, pagination, cancellation, and result envelopes live there.
3. Browsing is lazy and server-bounded. The client never presents a partially loaded client-side filter or sort as complete.
4. Relational rows, Redis values, MongoDB documents, search documents, and ClickHouse results retain model-native workflows.
5. Destructive work starts with a preview. Execution requires a short-lived confirmation token bound to the exact target and effect.
6. Credentials remain in Keychain. Saved connection files contain references and non-secret settings only.
7. Unsupported actions are hidden or disabled with a capability reason. They are never simulated through an unsafe fallback.
8. Network loss, cancellation, and a failed connection remain isolated from every other open session.

Workbench areas

The Database section contains:

• A searchable connection list with groups, tags, favorites, recent use, health, read-only state, and production labels.
• A lazy object explorer with breadcrumbs, favorites, scope refresh, permission states, and direct navigation between related objects.
• A tabbed workspace for data, query, command, definition, plan, monitoring, import, export, and visualization tasks.
• A bounded dynamic grid with server pagination, projection, filtering, sorting, selection, staged edits, detail inspection, and accessible keyboard behavior.
• Model-specific workspaces for Redis and Valkey keyspaces, MongoDB documents and pipelines, Elasticsearch and OpenSearch indices, and ClickHouse analytical structures.
• An operation center for progress, cancellation, partial failures, history, and safe retry guidance.

At wide widths the explorer, workspace, and inspector can be visible together. At narrower widths they become resizable focused columns, sheets, or tabs. Important actions move into overflow menus instead of disappearing. Grid columns scroll within the grid, while identifiers can remain pinned and full records remain available in a detail view.

Documentation map

• Research records inspected products and the design lessons adopted or rejected.
• Capabilities defines the supported feature surface and honest exclusions.
• Architecture defines contracts, data flow, adapters, safety, CLI, and MCP.
• Experience defines workflows, responsive behavior, accessibility, and visualization.
• Verification defines the real-product, large-data, failure, performance, and automation strategy.
• Delivery defines the pull request stack and completion gates.

Resource limits

Defaults are deliberately conservative and configurable within hard safety limits:

Resource	Default	Hard limit or behavior
Data page	200 records	2,000 records
Projected fields	Visible plus identity	512 fields
Large text preview	64 KiB	Explicit inspector load up to 8 MiB
Binary preview	Metadata only	Explicit chunked load up to 8 MiB
Metadata page	200 objects	1,000 objects
Schema inference	1,000 documents	10,000 documents
Redis scan count hint	500 keys	5,000 keys per request
Export batch	2,000 records	Adaptive, memory-bounded
Import batch	1,000 records	Adapter and server bounded
MCP result page	100 records	500 records
Operation history	500 entries	Oldest completed entries evicted
Concurrent connections	8	32 with per-connection pools
Query timeout	30 seconds	Explicit override required

Counts are estimated by default when exact counting would require a full scan. Continuation tokens are opaque, scoped, expiring, and invalid after a connection or query definition changes.

Safety summary

Saved connections can be read-only, mutation-enabled, or production-protected. Read-only enforcement occurs in the command executor and adapter, not only in the interface. Production mutations require a preview even for a single record. Unrestricted updates and deletes, truncation, drops, permission changes, session termination, ClickHouse mutations, and search delete-by-query use stronger confirmation language.

Preview responses contain the connection, environment, product, namespace, target, predicate or selection, normalized statement or request, estimated effect, transaction behavior, rollback availability, synchronous or asynchronous behavior, warnings, and an expiring token. The token is signed locally and binds those fields. Editing any field invalidates it.

The executor never silently retries a non-idempotent operation. Logs, errors, history, structured output, MCP results, screenshots, and evidence pass through the same secret redactor.