Miudb

miudb

Headless database CLI for agent-safe SQL work against saved database connections. Prefer direct miudb commands for one-shot machine-readable output and scripted checks. Use miudb mcp serve only when configuring an MCP host, and use miudb serve only for Neovim/custom protocol clients.

Do not use sqlit for miudb tasks.

When to use

• "Run this SQL on ``"
• "List schemas/tables/columns for ``"
• "Smoke-test my saved database connections"
• "Check a tunnel-backed connection"
• "Add a new database connection"
• Any task naming a connection that appears in miudb connections list

Install/verify

brew install vanducng/tap/miudb   # first install
miudb upgrade                      # self-update to the latest release
miudb version --output json        # v0.8.0 or newer
miudb commands --output json

If the catalog lacks a command you need, run miudb upgrade, or build from the local checkout:

cd           # your local clone of the miu-db source repo
go build -buildvcs=false -o ./.miu-db/miudb ./cmd/miudb

Default local config

miudb uses a native Go store by default:

~/.config/miu/db/connections.json
~/.config/miu/db/credentials.json

Sensitive values are classified before persistence. New database and SSH passwords are stored outside connections.json by default using the OS Keychain/keyring service named miudb.

For migrated configs, miudb reads credentials-export.json from the same directory when credentials.json is absent.

Only pass --config-dir, --connections-file, --credentials-file, --credentials-export, --secret-source, --keyring-service, or --gopass-prefix when the user asks for a non-default store.

By default, secret lookup can use file credentials, the miudb keyring service, and gopass paths under the miudb prefix. Treat --credentials-export as a deprecated alias for --credentials-file.

Discover commands and connections

miudb commands --output json
miudb describe connections add --output json
miudb describe connections test --output json
miudb describe query run --output json
miudb describe connections smoke --output json
miudb describe mcp serve --output json
miudb connections list --output json
miudb connections list --basic --output json   # scannable: ref/name/group/db_type/host

Connections are addressed by group/name (e.g. cnb/cdljn-prod); a bare name works when it is unique across groups. The ref column from --basic is exactly what to pass to --connection/-c. If the named connection is not listed, stop and ask the user. Do not substitute a similar connection.

Add connections

SQLite:

miudb connections add \
  --name local-app \
  --db-type sqlite \
  --path ./app.db \
  --output json

Postgres/MySQL style TCP connection:

miudb connections add \
  --name app-dev \
  --db-type postgresql \
  --host localhost \
  --port 5432 \
  --database app \
  --username app \
  --password "$APP_DB_PASSWORD" \
  --secret-store keyring \
  --output json

Tunnel-backed connection:

miudb connections add \
  --name app-prod \
  --db-type mysql \
  --host prod-rds.internal \
  --port 3306 \
  --database app \
  --username app \
  --password "$APP_DB_PASSWORD" \
  --tunnel \
  --ssh-config-alias bastion \
  --secret-store keyring \
  --output json

Provider options:

miudb connections add \
  --name warehouse \
  --db-type snowflake \
  --host account \
  --username USER \
  --option authenticator=snowflake_jwt \
  --option warehouse=DEV_WH \
  --extra-option sslmode=require \
  --output json

Secret stores for new connections:

• keyring: OS Keychain/keyring service named miudb on CGO-enabled builds.

Release binaries (brew install) are CGO-disabled and fall back to an encrypted file keyring at ~/.config/miu/db/keyring/: (JWE / PBES2-HS256+A128KW, one file per secret) — NOT the OS Keychain, so security find-generic-password -s miudb ... finds nothing there.

• file: local credentials.json with mode 0600.
• inline: leave the value in connections.json.
• none: discard the supplied secret and require another resolver later.

Rules:

• Prefer --secret-store keyring for user-entered credentials.
• Use --password-command only when the command is already trusted by the

user; never invent a credential command.

• Use --secret-store file only for disposable/local test configs or when the

user explicitly wants a file-backed credential.

• Never print passwords, credential files, private keys, or service account

JSON.

Verify where a secret landed (cheapest first, never reveals the value):

• miudb connections list --output json -> has_password: true, and the

connection's secrets[] shows provider: keyring + ref.

• ls ~/.config/miu/db/keyring/ -> :db file on release builds

(contents are JWE-encrypted; the raw password is not in the file).

• miudb connections test -> ok: true proves it resolves end-to-end.

Test one connection

miudb connections test  \
  --timeout 12s \
  --output json

Use connections test when the user names one connection or asks whether one connection is reachable. This opens the connection and may create an SSH tunnel, but does not run user SQL.

OAuth login for Snowflake and BigQuery

Acquire and store an OAuth token for a connection supporting OAuth (Snowflake, BigQuery):

miudb auth login  --output json
miudb auth status  --output json
miudb auth logout  --output json

Tokens are stored in the keyring service. On CGO-disabled builds (release binaries), fallback to file-based credential storage.

Smoke-test connections

miudb connections smoke \
  --timeout 12s \
  --concurrency 4 \
  --output json

Interpretation:

• Top-level ok: false can be expected when local-only databases are stopped.
• Check .data.results[] for per-connection pass/fail.
• Local failures like localhost:3307 refused usually mean a local service or

tunnel is not running.

• Tunnel failures can mean SSH alias/key/remote network issues.

Summarize results without printing passwords or secret file contents.

Run a query

miudb query run \
  --connection  \
  --sql '' \
  --limit 100 \
  --output json

Rules:

• Keep --limit bounded unless the user explicitly asks for a large export.
• Prefer read-only SQL unless the user explicitly authorizes mutation.
• Use single quotes around SQL containing BigQuery/MySQL backticks.
• If SQL contains single-quoted literals and backticks, escape carefully; do

not assume file input exists unless miudb describe query run says it does.

Fetch paged results

If query run returns a cursor or truncation marker, continue with:

miudb query fetch-page \
  --cursor  \
  --output json

Run multi-statement scripts

For Snowflake and MySQL, execute multiple statements in one script; each statement produces one result set:

miudb query script \
  --connection  \
  --sql 'SELECT 1; SELECT 2' \
  --output json

Postgres rejects multi-command scripts; run statements individually with query run.

Inspect schema

miudb schema tree \
  --connection  \
  --output json

Metadata SQL recipes also work through query run.

BigQuery

miudb query run \
  --connection  \
  --sql 'SELECT schema_name FROM INFORMATION_SCHEMA.SCHEMATA' \
  --output json

Use single quotes for BigQuery table references:

miudb query run \
  --connection  \
  --sql 'SELECT * FROM `dataset`.`table` LIMIT 10' \
  --limit 10 \
  --output json

MySQL

miudb query run --connection  --sql 'SHOW DATABASES' --output json
miudb query run --connection  --sql 'SHOW TABLES' --output json
miudb query run --connection  --sql 'DESCRIBE `table_name`' --output json

Postgres and Snowflake

miudb query run --connection  --sql 'SELECT table_schema, table_name FROM information_schema.tables WHERE table_type = ''BASE TABLE''' --output json
miudb query run --connection  --sql 'SELECT column_name, data_type, is_nullable FROM information_schema.columns WHERE table_name = ''table_name'' ORDER BY ordinal_position' --output json

SQLite

miudb query run --connection  --sql "SELECT name FROM sqlite_master WHERE type = 'table'" --output json
miudb query run --connection  --sql "PRAGMA table_info('table_name')" --output json

Generate an ERD (interactive diagram)

miudb erd turns a connection into a self-contained, offline interactive ER diagram (Cytoscape) plus a DBML export. Tier-1: MySQL, Postgres. Snowflake/BigQuery/DuckDB return a clean "unsupported" error.

# interactive offline index.html + schema.json + schema.dbml (default --format html)
miudb erd generate -c  --out-dir .diagrams/-erd --output json
# serve it; auto-opens the browser for an interactive terminal (--no-open to suppress);
# --from  renders an existing export with no DB
miudb erd serve -c  --output json

Short flags on every erd command: -c/--connection, -s/--schema, -m/--meta, -f/--format, -p/--port. A connection with no default database (e.g. a server-level DSN) errors with a clear "pass --schema" hint.

Two layers:

• Deterministic (miudb): introspect -> schema.json (the render source-of-truth) -> DBML + HTML. No LLM.
• Agentic (you): author meta.json to add colored domain groups + table descriptions. Without it, tables render as Framework/Other with no colors (erd generate emits a warning saying so).

Agentic polish recipe (how to make a good diagram)

1. Scaffold: miudb erd meta --stub --connection --out-dir — auto-detects framework tables (Laravel/Rails/Django/Prisma) and seeds blank groups/descriptions. The envelope data.next_step reminds you what to fill.
2. Read /schema.json (the IR: tables[] -> {pk, columns, fks, indexes, rows}). Cheap, high-signal inputs: FK topology (which tables connect), table-name prefixes, FK hub degree, and row counts. Migration filenames (ls database/migrations / db/migrate) are a strong signal too — you do NOT need to read their contents.
3. Edit meta.json:

• groups: cluster every non-framework table into 5-9 domains. Densely-FK-connected tables belong together; split by name prefix and responsibility (catalog vs apply vs analytics vs CMS vs auth). Each group = { "color": "#hex", "tables": [...] }. Put core domains in saturated colors, infra/marketing in muted. Palette: #2563eb blue, #16a34a green, #9333ea purple, #d97706 amber, #dc2626 red, #0d9488 teal, #db2777 pink, #475569 slate.
• descriptions: 1 line per important table (hubs + biggest by rows), inferred from name + columns + FK role (self-ref FK -> "conditional/nested"; double-FK + unique pair -> "junction"; *_id hub -> "owns/links X").
• leave framework_tables / audit_columns as detected.

1. Regenerate: miudb erd generate --connection --meta /meta.json (or erd serve). Iterate.

Single-pass works: stub -> fill -> generate. Aim to leave 0 tables ungrouped (the renderer buckets ungrouped non-framework tables as "Other").

Worked example (a ~100-table SaaS schema): the stub detects the framework tables; the FK hubs (a high-degree users/accounts table, a few central domain tables) plus name prefixes map cleanly to ~6-8 domains (e.g. Catalog, Orders, Billing, Analytics, Content, Auth) — grouping every non-framework table with colors + hub descriptions in one pass. Use generic examples; never paste real connection/schema names into the diagram metadata you commit.

Viewer (the rendered index.html): click a table to spotlight its FK chain; Focus (View options) hides everything except the selected table + its relations — best for reading a dense schema; DBML (toolbar) shows the dbdiagram.io/dbdocs source with Copy; Domain Groups has all/clear; the header shows live visible/total counts; hover truncated names for the full value. Initial zoom is clamped so columns stay legible — the Fit button zooms to the whole graph. State (filters/selection) is encoded in the share URL.

Stdio protocol

For Neovim or client integration, use the experimental stdio server:

miudb serve --protocol jsonrpc --output json

Use this only for client/protocol tasks. For normal agent work, call the direct CLI commands above.

MCP server

For MCP-native hosts such as Codex, Claude Code, Cursor, and VS Code, use:

miudb mcp serve --transport stdio

Useful flags:

• --connection : repeat to restrict visible/callable connections.
• --limit : default row limit for MCP query tools.
• --max-limit : maximum accepted MCP query limit.
• --max-bytes : maximum serialized bytes per tool/resource response.
• --allow-mutate: allow mutation SQL through MCP query_run; unsafe.

MCP tools exposed by the server:

• connections_list
• connection_describe
• connection_test
• connections_smoke
• schema_tree
• query_run
• query_fetch_page

MCP query_run is read-only by default and rejects mutation SQL unless --allow-mutate is explicitly provided. Stdout is reserved for MCP frames; startup errors and diagnostics go to stderr.

Query activity log

Each session captures activity events in a per-session JSONL log. Query and prune it with:

miudb activity --connection  --since 24h --output json
miudb activity --failed --since 7d --output json
miudb activity prune --older-than 30d --output json

Use --since to filter by relative duration (e.g. 24h, 7d); omit to read all captured events. Use --failed to show only failed queries.

Output contract

• stdout is JSON.
• stderr is diagnostics only.
• ok: false is a structured failure, not necessarily a shell failure.
• Command descriptions are available via miudb describe .
• Output is secret-hardened: credential-named values, password-bearing URLs, and

key=secret assignments are redacted before stdout (connections list shows has_password: true, never the value). Query-result values are NOT masked — they're the user's data. Do not inspect credential stores unless asked.

• The command catalog currently includes connections test, mcp serve, and

the native serve protocol; choose the narrowest command that matches the user's use case.

Failure modes

• connection not found -> run connections list, then ask the user.
• localhost refused -> local database/tunnel is not running.
• secret timeout -> keyring/gopass lookup may need user session access.
• SSH/tunnel error -> check ~/.ssh/config, key path, username, and network.
• BigQuery auth error -> verify options.bigquery_credentials_path.
• Snowflake JWT error -> verify options.private_key_file.
• query too large -> lower --limit or ask before exporting.

Anti-patterns

• Do not run Python miu-db TUI for agent tasks.
• Do not use sqlit.
• Do not guess connection names.
• Do not print credentials, private keys, or service account JSON.
• Do not run destructive SQL without explicit user approval.
• Do not omit --output json for agent-consumed results.
• Do not use unbounded queries in conversation context.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vanducng
• Source: vanducng/skills
• License: MIT
• Homepage: https://skills.vanducng.dev

Install and usage instructions live in the source repository linked above.