Custom Mcp Database

Custom MCP Database

mcp-name: io.github.renanlido/custom-mcp-database

An MCP server that lets AI agents run alias-based queries against PostgreSQL, MySQL, MongoDB and Oracle — without ever exposing credentials to the model. Connections are configured once and stored locally; the agent only ever references them by alias.

Works with Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Gemini CLI, and any other MCP client (all use the same stdio launch command).

Quickstart

There are two roles, on purpose. Keeping them separate is what stops your DB password from ever reaching the model.

You (once, in your terminal) — install the credentials

The agent never installs credentials. You do, with the CLI. The secret stays on your machine and is never sent to the model.

Easiest way — the guided wizard (asks type, host, user, and how to supply the secret; optionally tests the connection):

uvx custom-mcp-database setup

Or do it in one line (you'll be prompted for the password — hidden input):

uvx custom-mcp-database add-db --alias prod_ro --type postgres \  --host db.internal --port 5432 --user reporting --dbname app
uvx custom-mcp-database list-aliases   # confirm it's there

The agent (always) — uses it by alias

Point your MCP client at the server (see Install), then just ask:

"Using prod_ro, run SELECT count(*) FROM orders."

The agent calls db_execute_query with the alias prod_ro — never a host, user, or password. It physically cannot see the credentials; they live in your local config, resolved only inside the server process at query time.

Why the agent can't add the DB: an MCP tool's arguments are produced and read by the LLM. If the agent typed your password into an add tool, that password would land in the model's context, the provider, and the logs. So credential setup is a human/CLI step by design. (Need an agent to wire connections in an automated pipeline? See MCP_DB_ALLOW_ADMIN_TOOLS in SECURITY.md — even then it only accepts a reference to a secret, e.g. an env-var name, never the secret itself.)

Writes are off by default (read-only). To allow them for a task: export MCP_DB_READONLY=0 MCP_DB_ALLOW_WRITES=1.

Install

The server runs over stdio. The universal launch command is uvx custom-mcp-database run (requires uv; the package is fetched from PyPI on first run).

Claude Code

# Direct (published package)claude mcp add custom-mcp-database -- uvx custom-mcp-database run
# Or install the full plugin from this repo's marketplace/plugin marketplace add renanlido/custom-mcp-database/plugin install custom-mcp-database@renanlido-mcp

Claude Desktop

Two options:

1. One-click bundle — build the .mcpb (mcpb pack) and open it in Claude Desktop. See Distribution.
2. Manual config — add the snippet from examples/mcp-clients/claude-desktop.json to claude_desktop_config.json.

Other clients

Copy the matching snippet — all use the same command/args, only the file and key differ:

Full client matrix and a local-checkout variant: examples/mcp-clients/README.md.

Configure connections

Configure connections from your terminal with the CLI — never through the agent. A connection's password is a real secret; if it were passed as an MCP tool argument it would enter the model's context (and the provider, transcripts, and logs). So the credential-management tools are off the MCP surface by default; provisioning is a human/CLI task. The agent only lists and uses aliases.

Omit --password/--uri to be prompted securely (hidden input, not stored in shell history). Even better, keep the secret out of the config file entirely with --password-env / --password-file (resolved at connection time):

# PostgreSQL — prompted for the password (recommended)uvx custom-mcp-database add-db --alias pg --type postgres \  --host localhost --port 5432 --user me --dbname app
# MySQL — password taken from an env var at connect time (nothing secret on disk)MYSQL_PW=... uvx custom-mcp-database add-db --alias my --type mysql \  --host localhost --port 3306 --user root --dbname app --password-env MYSQL_PW
# Oracle — password read from a file (e.g. a mounted secret)uvx custom-mcp-database add-db --alias ora --type oracle \  --host db.example.com --port 1521 --user system --dbname ORCLPDB1 \  --password-file /run/secrets/ora_pw
# MongoDB — full URI from a file (the URI embeds credentials)uvx custom-mcp-database add-db --alias mongo --type mongo \  --dbname app --uri-file /run/secrets/mongo_uri
uvx custom-mcp-database list-aliasesuvx custom-mcp-database remove-db --alias pg

Config location (override with MCP_DB_CONFIG): $XDG_CONFIG_HOME/custom-mcp-database/mcp_config.sqlite3 (default ~/.config/custom-mcp-database/mcp_config.sqlite3, 0600).

If you pass a literal --password/--uri, it is stored as plaintext JSON in that SQLite file. Prefer --password-env/--password-file (or --uri-env/--uri-file) so only a reference is stored. Either way, keep the file secret (it is 0600, gitignored, not encrypted).

MCP tools

db_add_database / db_remove_database are not exposed over MCP by default — manage connections with the CLI. To opt into exposing them (the add tool only accepts secrets by reference, never a literal password), set MCP_DB_ALLOW_ADMIN_TOOLS=1.

db_execute_query notes: SQL runs as given with parameterized binds (add your own LIMIT); MongoDB takes a JSON filter + collection, caps results at 10 (--limit), rejects empty filters, and coerces 24-char hex strings to ObjectId.

Security

This server handles real credentials and production data, so it ships deny-by-default:

• Read-only by default. Only SELECT-class SQL runs. Writes/DDL require explicit opt-in.
• No stacked statements (;-injection blocked), single statement per call.
• MongoDB server-side JavaScript blocked ($where, $function, $accumulator, mapReduce, …).
• Identifiers validated (oracle_schema can't be used for injection).
• Results capped at MCP_DB_MAX_ROWS (default 1000); secrets redacted from errors.
• Credential store is 0600 plaintext SQLite — keep the host disk encrypted.

Check the live posture: custom-mcp-database security-status (or the db_security_status tool).

Enable writes for a specific task (then turn it back off):

export MCP_DB_READONLY=0export MCP_DB_ALLOW_WRITES=1     # INSERT/UPDATE/DELETE# export MCP_DB_ALLOW_DDL=1      # only if you really need CREATE/DROP/ALTER/...

Read the full protocol — least-privilege DB roles, TLS, prompt-injection handling, vulnerability reporting — in SECURITY.md. The app-layer guards are defense-in-depth; the authoritative control is a least-privilege database account.

Develop

uv sync                 # create .venv and install depsmake run                # run the server (stdio)make lint               # ruffmake build              # sdist + wheel into dist/

Inspect tools interactively:

uv run mcp dev src/custom_mcp_database/server.py

Distribution

This repo ships ready-to-publish metadata for every major channel. All of it is published automatically on push to main (see below):

Automated release — just push to main

Releases are fully automated. On every push to main, .github/workflows/release.yml:

1. Picks the next semantic version from your commits since the last tag (feat: → minor, BREAKING CHANGE/type!: → major, anything else → patch; add [skip release] to a commit message to skip).
2. Writes that version into pyproject.toml and syncs it into every artifact (server.json, manifest.json, plugin + marketplace) via scripts/sync_version.py — version lives in one place, no hand-bumping.
3. Builds, commits chore(release): vX [skip ci], tags vX, pushes.
4. Publishes to PyPI (Trusted Publishing/OIDC), then the MCP Registry (GitHub OIDC).
5. Packs the .mcpb and cuts a GitHub Release with the wheel + bundle attached.

The release commit carries [skip ci], so it does not re-trigger the workflow.

One-time setup (can't be automated — needs your accounts):

• Create a PyPI Trusted Publisher for renanlido/custom-mcp-database, workflow release.yml.
• Allow GitHub Actions to push to main (repo → Settings → Actions → Read and write permissions; if main is a protected branch, allow the actions bot to bypass or use a PAT).

The MCP Registry namespace is io.github.renanlido/custom-mcp-database (GitHub-validated).

Local manual escape hatch: make build (syncs version + builds) then uv publish.

License

MIT