Liquid

What ertad-family/liquid MCP server does

The ertad-family/liquid MCP server provides a common interface for reaching different kinds of systems through four core operations: fetch, query, write, and sense. It can work with REST and GraphQL APIs, SOAP/WSDL, gRPC, WebSocket, SSE/NDJSON, MQTT, email systems, databases, other MCP or A2A agents, and several industrial or device protocols described by the project.

The server accepts an endpoint such as an HTTPS URL, a database DSN, a gRPC target, or another MCP server. It discovers the available interface and returns typed records shaped around the fields requested by the caller. Supported database examples include PostgreSQL, MySQL/MariaDB, SQLite, DuckDB, SQL Server, Neo4j, MongoDB, and Redis.

How it works

Discovery happens when an adapter is created. Interfaces with machine-readable descriptions, including OpenAPI, GraphQL, gRPC reflection, MCP, A2A, and WSDL, can be discovered without an LLM. Databases are introspected directly. An LLM backend is needed for undocumented REST APIs and field mapping; the README identifies LiteLLM as the route to OpenAI, Gemini, Anthropic, local models, and other providers.

After setup, the runtime transport is deterministic and does not make model calls for each operation. Liquid can flatten common pagination envelopes, normalize identifiers and timestamps, and normalize certain cross-source values such as monetary fields. It also supports server-side search, aggregation, text search, token limits, and verbosity controls. Errors can include a structured recovery action, such as retrying after rate limiting or repairing an adapter after an upstream change.

Setup and configuration

Install the package with pip install liquid-api. The bundled MCP command can be started with uvx --from 'liquid-api[mcp]' liquid-mcp. The project documents Python 3.12 or newer. Add the discovery extra when the server must use an LLM to discover undocumented REST APIs or map fields. Optional extras are documented for Gemini, gRPC, WebSocket, PostgreSQL, MySQL/MariaDB, Neo4j, and DuckDB.

The upstream connection is supplied as a URL, DSN, or protocol target. Credentials are handled as part of the upstream connection process, but the provided material does not define a single required environment variable or authentication configuration for the MCP command.

Tools and capabilities

The ertad-family/liquid MCP server exposes capabilities corresponding to these operations and supporting functions:

• Fetch typed records from an adapter.
• Query, search, aggregate, or text-search upstream data.
• Write data when write access is explicitly enabled.
• Sense live events from streams, mailboxes, MQTT, devices, or supported systems.
• Estimate expected output size before fetching.
• Normalize fields across sources and repair adapters after schema drift.
• Create tool metadata describing cost, latency, caching, idempotency, and side effects.

The project also describes canonical intents for some actions and ambient tools for checking quota or listing adapters. Exact availability can depend on the selected transport and installed extras.