Pack de Connecteurs MCP pour Bases de Données

What's in this pack

This is the broader database MCP pack. Where Postgres for AI Agents goes deep on one engine, this one cuts across all five databases an average backend touches — Postgres, MySQL, MongoDB, Redis, SQLite — plus two universal routers and a credential-free dry-run server for the truly paranoid.

Five engines, three safety levels (read-only, capability-scoped, offline), one install command.

Why a multi-database MCP pack matters

Most teams don't run a single database. The agent that ships features against the user-facing Postgres also needs to read the analytics MySQL, the session Redis, the catalog MongoDB, and the local SQLite the developer is testing against. Wiring each engine into the agent host with its own ad-hoc MCP server is how you end up with five servers, three credential formats, and zero consistent audit trail.

This pack picks the connectors that get the same things right across engines:

• Read-only mode is the default. Every server here exposes a --read-only or equivalent capability flag. Write access is opt-in, not opt-out.
• Schema introspection is a first-class tool. The agent calls list_tables, describe_table, list_indexes instead of guessing.
• Connection pooling is built in. No reconnect storms when the agent fires twenty queries during a single reasoning step.
• Audit logging is on by default. Every query the agent runs lands in a structured log you can grep after an incident.

The outlier is DryRun (#3): it doesn't touch your database at all. You hand it a schema dump and it lets the agent draft SQL, get type-checked, and explain the plan — all without credentials. Useful when the agent is helping a junior engineer write a migration before anyone's allowed near prod.

Install — pick DB, set read-only, introspect, query, audit

The five-step rhythm that every server in this pack supports:

# 1. Pick the DB you want the agent to see
tokrepo install pack/mcp-database-connectors
# or one at a time:
tokrepo install postgres-mcp-pro
tokrepo install mongodb-mcp-server

# 2. Wire it into your host with read-only ON
# Example for Claude Code (~/.claude/mcp.json):
{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@pgEdge/pg-mcp", "--read-only"],
      "env": { "DATABASE_URL": "$READONLY_DSN" }
    }
  }
}

# 3. Ask the agent to introspect FIRST, query SECOND
#   Prompt: "list the tables, describe the orders table,
#            then write a query for last week's revenue."

# 4. Query — agent runs the SELECT, you read the result

# 5. Audit — grep the server log
tail -f ~/.tokrepo/logs/postgres-mcp-pro.log

For MySQL swap in mcp-server-mysql, for Mongo swap in mongodb-mcp-server, for Redis swap in the official redis-mcp-server. The five steps are identical — the pack's value is that you only learn them once.

Read-only mode: why every flag matters

A naive write-enabled MCP is the fastest way to turn an LLM hallucination into a production outage. The classic failure mode: agent reasons about an UPDATE, gets the WHERE clause wrong, and rewrites half the table.

Every server in this pack flips this default. The agent has to explicitly request write capability, and your host config has to explicitly grant it. That's a two-step opt-in, which is exactly what you want for anything that can DROP TABLE.

When you do need writes — running a seeded migration, populating a test fixture — use a dedicated role (mcp_writer) with grants limited to the specific schemas, never the superuser. Combine with a branch (Neon), a clone (pgEdge), or a Docker fixture (everything else) and the blast radius collapses to one throwaway database.

Common pitfalls

• Schema dump cost. list_tables on a 500-table production schema can blow the agent's context window. Filter to a single schema or pass an allow-list.
• The universal DB MCP isn't free of engine quirks. It handles 17 databases, but the JSON-column semantics differ between Postgres and MySQL — the agent will sometimes write Postgres-flavored SQL against MySQL. Add a hint to the system prompt.
• Redis ACLs are per-command, not per-key. The official server respects ACLs, but a KEYS * is still O(N) — set --max-scan-keys or the agent will block the event loop.
• MongoDB's --read-only blocks writes but not eval. Older Mongo MCPs let the agent run server-side JavaScript. Stick to the post-2026 builds in this pack which strip $where and $function.
• DryRun's schema must stay fresh. It's offline by design, so it can't notice a column you added yesterday. Re-export the schema as part of CI.

How this pack relates to the others

If you only care about Postgres in depth (Neon, Supabase, branching), use Postgres for AI Agents. If you want the broader MCP surface (browser, GitHub, filesystem), use MCP Server Stack. Production setups usually run one item from each: a database connector here, the universal MCP toolbox there, and a vector DB for semantic recall.