CoreMCP
About
Connect Legacy Databases to AI Agents via Model Context Protocol. Open-source bridge for LLM data analysis.
Details
- Author
- corebasehq
- Categories
- Database, Other, AI
Jump to
Setup
Install CoreMCP in your MCP client (Claude Desktop, Cursor, Windsurf, and others).
Repository: https://github.com/corebasehq/coremcp
Follow the installation instructions in the repository README, then restart your MCP client.
A Model Context Protocol (MCP) server, written in Go, that exposes SQL databases as MCP tools and prompts. It runs as a single static binary, embeds its drivers, and talks either stdio (for local MCP clients like Claude Desktop) or an outbound WebSocket (for remote operation behind NAT).
Currently ships with MSSQL (SQL Server 2000+,Turkish_CI_AScollation aware) and PostgreSQL adapters. Firebird is in progress; MySQL is on the roadmap.
- Stable:MSSQL adapter, PostgreSQL adapter, stdio transport, schema discovery, custom tools, NOLOCK / Turkish normalization middleware, WebSocket connect mode.
- In progress:Firebird adapter (factory currently returns a placeholder error).
- Roadmap:MySQL, HTTP transport, audit log, query result cache.
CoreMCP is read-only by default. Omittingreadonlyin a source config leaves SELECT-only mode active; you have to setreadonly: falseto enableexecute_procedure. Even so, the recommended posture is a dedicated DB user withSELECT(andEXECUTEonly on the procedures you intend to expose) — defense in depth rather than relying solely on the server-side guard.
Download from theReleases page—linux/amd64,linux/arm64,darwin/{amd64,arm64},windows/amd64.
curl -fsSL https://get.corebasehq.com | sh
Multi-arch image (linux/amd64,linux/arm64).
git clone https://github.com/corebasehq/coremcp.git cd coremcp go build -o coremcp ./cmd/coremcp
server: name: "coremcp-agent" version: "0.1.0" transport: "stdio" port: 8080 logging: level: "info" format: "json" sources: - name: "my_database" type: "mssql" dsn: "sqlserver://username:password@localhost:1433?database=mydb&encrypt=disable" readonly: true no_lock: true # READ UNCOMMITTED isolation (WITH (NOLOCK) equivalent) normalize_turkish: true # Turkish character + mojibake normalization
Seecoremcp.example.yamlfor a fuller example.
sqlserver://username:password@host:port?database=dbname&encrypt=disable
postgresql://username:password@host:port/dbname?sslmode=disable
Dummy adapter (for testing without a real DB):
sources: - name: "oltp_db" type: "mssql" dsn: "sqlserver://user:pass@localhost:1433?database=production&encrypt=disable" readonly: true no_lock: true
sources: - name: "erp_db" type: "mssql" dsn: "sqlserver://user:pass@localhost:1433?database=LOGO&encrypt=disable" readonly: true no_lock: true normalize_turkish: true
security: max_row_limit: 1000 # forced LIMIT cap enable_pii_masking: true pii_patterns: - name: "credit_card" pattern: '\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b' replacement: "*---" enabled: true - name: "email" pattern: '\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b' replacement: "@." enabled: true - name: "turkish_id" pattern: '\b[1-9]\d{10}\b' replacement: "" enabled: true
- T-SQL aware lexer.Fail-closed custom tokeniser strips comments and string literals, then classifies the statement — onlySELECTandWITHpass.DROP,ALTER,UPDATE,DELETE,TRUNCATE,EXEC,OPENROWSET,SELECT…INTOand similar are rejected before reaching the DB. Multi-statement payloads (any;outside strings/comments) are fatal — stacked-query attacks blocked dialect-independently. Chosen over third-party Go SQL parsers (xwb1989/sqlparser, vitess, cockroachdb) because they fail-closed on T-SQL hints and any "fall through to regex" relaxation is bypassable viaEX//ECand similar tricks. Treat as one layer, not the only layer — pair with a least-privilege DB role.
- Forced row cap.LIMITis appended (or wrapped) on every SELECT so a model never streams millions of rows back through the protocol.
- PII masking.Regex-based post-processing on result strings before they reach the client.
For local MCP clients (Claude Desktop, etc.):
Claude Desktop config (claude_desktop_config.json):
{ "mcpServers": { "coremcp": { "command": "/path/to/coremcp", "args": ["serve", "-c", "/path/to/coremcp.yaml"], "env": {} } } }
connectopens an outbound WebSocket to a relay (typically CoreBase Cloud) and serves MCP traffic over it. The agent never accepts inbound connections, so it works from inside networks that don't allow inbound 443 (factory floors, corporate VPCs, hospital networks).
coremcp connect --server="wss://api.corebasehq.com/ws/agent" --token="sk_xxx"
-s, --server string Relay WebSocket URL (required) -t, --token string Authentication token (required) -a, --agent-id string Agent ID (auto-generated if omitted) -r, --max-reconnect int Max reconnect attempts (default 10; 0 = infinite) -d, --reconnect-delay duration Delay between reconnect attempts (default 5s)
./coremcp connect \ --server="wss://api.corebasehq.com/ws/agent" \ --token="sk_xxx" \ --agent-id="site-istanbul-001" \ --max-reconnect=0
Wire commands supported by the relay protocol:
- run_sql— execute SQL
- get_schema— dump cached schema
- list_sources— enumerate configured sources
- health_check— agent liveness
- config_sync— push updated source configs to the running agent
coremcp/ ├── cmd/coremcp/ # CLI entry point │ ├── main.go │ ├── root.go │ ├── serve.go # stdio mode │ └── connect.go # WebSocket mode ├── pkg/ │ ├── adapter/ # Database adapters │ │ ├── factory.go │ │ ├── dummy/ │ │ └── mssql/ │ ├── config/ │ ├── core/ # Shared types, Source interface │ ├── security/ # Query validation, PII masking │ └── server/ # MCP server └── coremcp.yaml
Arbitrary SQL against a configured source.
Tables with column counts, primary keys, foreign key counts.
Full schema for one table: columns, types, nullability, PKs, FKs, column comments.
- source_name(required)
- table_name(required)
Stored procedures with parameter names, types, modes (IN/OUT/INOUT), and a ready-to-copy example call.
Calls a stored procedure with named parameters.Only enabled whenreadonly: false.
- source_name(required)
- procedure_name(required)
- params(optional) — JSON object of name/value pairs
- Procedure name validated against^[a-zA-Z_][a-zA-Z0-9_#@.]$
- Parameter names validated (alphanumeric + underscore)
- Values bound viasql.Named— no string interpolation
- Rejected outright when source isreadonly: true
{ "source_name": "erp_db", "procedure_name": "sp_CiroHesapla", "params": "{\"StartDate\":\"2024-01-01\",\"EndDate\":\"2024-12-31\"}" }
Define reusable parameterized queries as first-class MCP tools:
custom_tools: - name: "get_daily_sales" description: "Daily sales summary for a given date" source: "production_db" query: "SELECT FROM orders WHERE DATE(created_at) = '{{date}}'" parameters: - name: "date" description: "Date in YYYY-MM-DD format" required: true - name: "get_top_customers" description: "Top N customers by order count" source: "production_db" query: "SELECT user_id, COUNT() AS order_count FROM orders GROUP BY user_id ORDER BY order_count DESC LIMIT {{limit}}" parameters: - name: "limit" description: "Number of customers to return" required: true default: "10"
These get exposed to the model with their declared parameter schema, so the model can call them directly rather than re-deriving the SQL each turn.
On startup CoreMCP connects to every configured source, scans tables / columns / keys / relationships, and extracts column comments (e.g.MS_Descriptionon MSSQL). The result is exposed as a single MCP prompt that primes the model with schema context — including the comments — so it can write correct queries without manual schema dumps in every conversation.
- Createpkg/adapter/yourdb/.
- Implementcore.Source.
- Register inpkg/adapter/factory.go.
pkg/adapter/dummy/dummy.gois the minimum reference implementation.
- Schema discovery on startup
- Column comments / descriptions
- Built-inlist_tables/describe_table
- Custom parameterized tools
- T-SQL aware lexer for query sanitization (fail-closed, multi-statement reject, no third-party parser)
- PII masking
- Forced row cap
- WebSocketconnectmode
- Auto-reconnect
- Remote config sync
- NOLOCK / READ UNCOMMITTED per source (MSSQL)
- Turkish character + mojibake middleware (MSSQL)
- View and procedure discovery (list_views,list_procedures,execute_procedure)
- PostgreSQL adapter
- Firebird adapter (in progress)
- MySQL adapter
- HTTP transport
- Query result cache
- Write operations (with explicit safety guards)
- Audit logging
- Multi-agent management
- Real-time monitoring
SeeCONTRIBUTING.md. Security reports:SECURITY.md.
- Report a bug
- Request a feature
- Email:support@corebasehq.com
CoreMCP is the open-source, on-prem gateway component ofCoreBase, an AI agent platform for your company's data. Chat with your databases and APIs directly, or let autonomous and event-triggered agents run multi-step automations across them — databases (SQL Server 2000+, PostgreSQL), REST and GraphQL APIs, and 50+ SaaS connectors.
CoreMCP is how those agents reach the systems behind your firewall, including the legacy and on-prem ones nothing else connects to: it runs on your own server, keeps database credentials local, and connects zero-trust — outbound port 443 only, no inbound ports. On top of that access, CoreBase layers Unified Context and Query Memory: the schema relationships, terminology, and proven query patterns that turn raw access into accurate answers.
AI-powered SQL query builder — natural language to SQL, schema introspection, query optimization, multi-dialect support by MEOK AI Labs
An MCP server that allows LLM agents to seamlessly execute functions within Unity Catalog.
Official MCP server for dbt (data build tool) providing integration with dbt Core/Cloud CLI, project metadata discovery, model information, and semantic layer querying capabilities.
Query and analyze data with MotherDuck and local DuckDB
An MCP server that provides tools to interact with Powerdrill datasets, enabling smart AI data analysis and insights.
A collection of tools for managing the platform, addressing data quality and reading and writing to Teradata Database.
Interact with AskTable SaaS or local deployments to query data sources using natural language.
A read-only MCP server for Avro data sources, powered by the CData JDBC Driver.
Run SQL queries on data in Amazon S3 using AWS Athena.
Interact with Bauplan data tables and run queries.
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
