Supports MariaDB via the MySQL/MariaDB adapter, enabling schema exploration, bounded SQL execution, schema snapshots/diffs, and governed change workflows with audit.
Provides MySQL database integration through the MySQL/MariaDB adapter, enabling schema exploration, bounded SQL execution, schema snapshots/diffs, and governed change workflows with audit.
Provides PostgreSQL database integration with schema exploration, dependency and procedure call-graph analysis, bounded SQL execution and explain plans, health/workload diagnostics and index recommendations, schema snapshots/diffs, and governed change workflows with audit.
Provides SQLite database integration for schema exploration, bounded SQL execution, schema snapshots/diffs, and governed change workflows with audit.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Self-hosted database MCP server written in Rust for coding agents (Claude Code, Cursor, Windsurf, Claude Desktop, VS Code, any MCP client).
It combines, in one binary:
Capability | What the agent gets |
Schema exploration |
|
Lineage and call graphs |
|
Bounded SQL |
|
Declarative tools |
|
PostgreSQL intelligence |
|
Governed changes |
|
Schema snapshots |
|
Audit | Structured JSONL + SQLite audit trail for every tool call, with secret redaction |
Engines: PostgreSQL, MySQL/MariaDB, SQLite. When a PostgreSQL or MySQL DSN is unreachable the adapter falls back to a deterministic embedded simulation so the full tool surface can be exercised offline (useful for CI and demos; see research/findings/capability-matrix.md).
Transports: stdio (local agents) and Streamable HTTP JSON-RPC (POST /mcp, Bearer auth, /health, /metrics, /admin/approve).
npx @vinoth_n_3000/db-mcp init # writes ./db-mcp.toml starter config
npx @vinoth_n_3000/db-mcp verify-config --config ./db-mcp.toml
npx @vinoth_n_3000/db-mcp serve --transport stdio --config ./db-mcp.tomlOn first run the launcher downloads the prebuilt db-mcp-server binary for your platform from the
GitHub release matching the package version, verifies its SHA-256,
and caches it under ~/.db-mcp/bin/. If no prebuilt binary exists for your platform it falls back to
cargo install from this repository (requires Rust), or to a binary you point at with DB_MCP_BINARY_PATH.
cargo install --git https://github.com/vinothhacks/db-mcp --locked db-mcp-server
db-mcp-server serve --transport stdio --config ./db-mcp.tomldocker compose up --build # db-mcp on :8080 plus disposable Postgres and MySQL
curl -s http://127.0.0.1:8080/healthgit clone https://github.com/vinothhacks/db-mcp && cd db-mcp
npm run build # cargo build --release -p db-mcp-server
npm test # stdio + HTTP + CLI approval end-to-end
npm run verify:features # exercises all 16 tools and prints a feature matrixRelated MCP server: PGScope MCP
The server speaks MCP over stdio. Point your client at the launcher (or the binary) and a config file.
Claude Code
claude mcp add db-mcp -- npx -y @vinoth_n_3000/db-mcp serve --transport stdio --config /abs/path/db-mcp.tomlCursor / Windsurf / Claude Desktop / VS Code (mcp.json, claude_desktop_config.json, .vscode/mcp.json)
{
"mcpServers": {
"db-mcp": {
"command": "npx",
"args": ["-y", "@vinoth_n_3000/db-mcp", "serve", "--transport", "stdio", "--config", "/abs/path/db-mcp.toml"],
"env": { "DATABASE_URL": "postgres://user:pass@host:5432/app" }
}
}
}More client recipes, HTTP mode and the admin approval flow: docs/CLIENT-SETUP.md.
config/db-mcp.example.toml is a complete, validated example. The important parts:
server_name = "db-mcp"
default_mode = "ReadOnly" # Development | ReadOnly | GovernedProduction
active_profiles = ["Explore", "Query", "Diagnose", "Change", "Schema"]
[http]
bind = "127.0.0.1:8080"
auth_token = "env:DB_MCP_AUTH_TOKEN" # Bearer token for POST /mcp (optional)
admin_token = "env:DB_MCP_ADMIN_TOKEN" # Bearer token for POST /admin/approve
[audit]
jsonl_path = "research/traces/audit.jsonl"
sqlite_path = "research/traces/audit.db"
[changes]
storage_path = "research/traces/changes.db"
policy_revision = "policy_v1"
default_ttl_seconds = 3600
[[connections]]
id = "prod-postgres"
engine = "Postgres" # Postgres | MySql | Sqlite
dsn = "env:DATABASE_URL" # literal DSN, env:VAR, or file:/path
mode = "GovernedProduction"
read_only = false
timeout_ms = 10000
max_rows = 500
max_bytes = 1048576
restricted_tables = ["internal_secrets"]
sensitive_columns = ["password_hash", "api_key", "secret_token"]
allowed_roles = ["agent", "developer", "dba", "admin"]
[[named_tools]]
name = "get_user_by_id"
connection = "prod-postgres"
statement = "SELECT id, email, role FROM users WHERE id = ?1"
read_only = true
[[named_tools.parameters]]
name = "user_id"
param_type = "integer"
required = trueValidate any config with db-mcp verify-config --config <file>. Secrets are never echoed: DSNs resolve through
env: / file: references and are redacted in logs, errors and audit records.
Mode | Reads | Writes / DDL | Notes |
| yes | direct | local fixtures, still bounded by limits |
| yes | denied ( | mutating CTEs and privileged functions blocked |
| yes | only via approved | self-approval, replay, hash tampering and schema drift are rejected |
agent: propose_change -> chg_123 (PENDING, approval_hash)
admin: db-mcp approve --change-id chg_123 --approver release-admin # CLI, or POST /admin/approve
agent: apply_change -> executes once, re-verifies hash + schema fingerprint, auditedExplore: list_connections, search_objects, describe_object, trace_table_dependencies, get_procedure_call_graph
Query: execute_query, explain_query, run_named_tool
Diagnose: database_health, analyze_workload, recommend_indexes
Change: propose_change, get_change, apply_change
Schema: capture_schema, diff_schema
Every success response carries the same envelope: request_id, connection, status, columns, rows,
returned_rows, truncated, warnings, schema_version, plus _meta.provenance.untrusted_content = true so
agents treat database content as data, not instructions. Denials are structured: status: "denied", code,
message, next_action. Full contracts: research/contracts/tool-contracts.json.
Call any tool from a shell for debugging:
node scripts/mcp-call.js --config config/db-mcp.example.toml trace_table_dependencies '{"connection":"local-sqlite","table_name":"users","target_table":"order_items"}'
node scripts/mcp-call.js --raw tools/listcrates/ 15 Rust crates (core, secrets, sql, catalog, policy, audit, executor, tools,
sqlite, postgres, mysql, pg-insights, changes, mcp, server)
npm/bin/db-mcp.js npx launcher: resolves, downloads or builds the native binary
npm/test/ Node end-to-end suite (stdio, HTTP, CLI approval)
scripts/ mcp-call.js (ad-hoc client), feature-smoke.js (all-feature verification)
config/ example configuration
docs/ client setup, feature verification report
research/ contracts, capability matrix, provenance, license inventory
tests/ contract, integration and adversarial-security suite indexes (run by cargo)cargo test --workspace # unit + contract + adversarial security + runtime integration
npm test # Node E2E over the npx launcher
npm run verify:features # feature matrix (used by CI)Dual-licensed under MIT or Apache-2.0, at your option. See LICENSE-MIT and LICENSE-APACHE.
Reference projects studied during design (DBHub, Postgres MCP Pro, Google MCP Toolbox, Bytebase) are documented in
research/licenses/license-inventory.md; DB-MCP is a clean-room Rust implementation.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Deterministic safety, correctness & cost gate that vets Postgres SQL before your AI agent runs it.
Safe, read-only Postgres and MySQL access for AI agents. Audit log + column-level controls.
Supabase-compatible Postgres for agents: projects, branches, SQL, auth, storage, email.
Query Postgres, MySQL, SQL Server, Oracle, BigQuery, ClickHouse and Redshift from your AI client.
