mnemiq
agenticfabriq.com
- Free plan
- Web
- API
- Self-hosted
What mnemiq is
mnemiq is an open-source text-to-SQL engine for enterprise data agents that can be configured and measured against a user's own database. Before execution, a deterministic checker verifies that proposed SQL is read-only, uses permitted objects, compiles in the database dialect, and passes EXPLAIN. Access permissions apply before schema retrieval, so the model is not shown tables the caller cannot access. Each answer includes the SQL run, the tables it touched, and the enrichment version used. The engine can profile columns and build descriptions, grain, glossary terms, and coded-value meanings; a human owner can certify semantic definitions. Listed database support includes PostgreSQL, SQLite, DuckDB, Oracle, Snowflake, and Databricks, with DuckDB as the universal executor. Models can use any OpenAI-compatible /v1 endpoint, including local servers. Its MCP server exposes two read-only, access-scoped tools, and one process can serve a workbench and HTTP endpoints. The Apache-2.0 engine is free and runs in the user's environment; managed grading, drift, and control-plane services are separate commercial offerings.
Who it suits
mnemiq suits teams building enterprise data agents that need to query their own databases with access controls and query traceability. It may also fit teams that want to run a local model server within their network.
Good to know, and what to watch
- Checks proposed SQL before execution.
- Applies permissions before schema retrieval.
- Answers include executed SQL and touched tables.
- Supports several database systems and model endpoints.
- Free Apache-2.0 engine runs in the user's environment.
- Managed grading and control-plane services are commercial.
- Repair-enabled benchmark averaged around 2.3 LLM calls per question.
- No mnemiq-specific price is stated for managed services.
Plans
Every plan and priceVerity and Agentic Fabriq control plane
mnemiq
Open source
The verdict
mnemiq combines guarded SQL execution with access-scoped schema retrieval and traceable answers. Teams should account for its reported LLM-call usage and distinguish the free engine from separate managed services.
What the maker publishes
- Access control
- Permissions are applied before schema retrieval, so the model is not shown tables the caller may not access. source
- Agent integration
- The MCP server exposes two read-only, access-scoped tools: db_read(question) and get_schema(). source
- API and workbench
- One process can serve a workbench and HTTP endpoints POST /v1/ask, POST /v1/chat, and GET /v1/schema. source
- Commercial services
- The maker identifies Verity, a managed grading and drift service, and the Agentic Fabriq control plane as commercial services around the open-source engine. source
- Company
- The mnemiq page identifies Agentic Fabriq as its maintainer; the opened maker pages did not state a headquarters or founding year. source
- Cost consideration
- The launch article says mnemiq averaged around 2.3 LLM calls per question with repair enabled in its benchmark comparison, while the compared vendor APIs were called once. source
- Database support
- The page lists PostgreSQL, SQLite, DuckDB, Oracle, Snowflake, and Databricks, with DuckDB as the universal executor. source
- LLM support
- Any OpenAI-compatible /v1 endpoint can serve the model, including local servers such as vLLM, Ollama, llama.cpp server, and LM Studio. source
- Purpose
- mnemiq is an open-source text-to-SQL engine for enterprise data agents that can be configured and measured on a user's own database. source
- Query checks
- A deterministic decider checks that proposed SQL is read-only, uses allowed objects, compiles in the source dialect, and passes EXPLAIN before execution. source
- Security deployment
- The maker says that using a local model server keeps every schema, question, and row inside the user's network. source
- Semantic enrichment
- It can profile columns and build descriptions, grain, glossary terms, and coded-value meanings, with semantic definitions optionally certified by a human owner. source
- Traceability
- Each answer includes the SQL that ran, the tables it touched, and the enrichment version used. source
Checked against the maker's pages on 4 Oct 2026.