Skip to main content

Your first query

A 5-minute walkthrough for a new engineer whose org has already deployed the gateway. Start to finish: install Claude Code → sign in → run a real query against a real DB. There's an optional last step for operators who want to peek at the audit row directly — it needs psql and state-DB access, so most engineers can stop after step 5.

If you're standing the gateway up, you want ../deployment/quickstart.md instead. If you've already done a first query and want the day-to-day reference, jump back to claude-code.md. For MongoDB targets the path is the same — see multi-db.md for what changes at the sql field and which commands are allowed.

Placeholder: every https://db.internal.acme.com below is a stand-in. Your platform team will give you the real one (looks the same, different domain).

Before you start​

You needHow to check
Claude Code installed and logged in to your Anthropic accountclaude --version
Your gateway URLAsk your platform team — it'll be on the Discord pin or the internal wiki
An SSO account at your org's IdPThe same login you use for the rest of your internal tools
Membership in at least one group with a grantWithout this you authenticate fine but list_servers returns nothing — see Gotchas

That's it. You don't need: a DB password, a VPN profile (unless your gateway is private), or any local DB driver. The optional audit-row peek in step 6 needs psql plus operator-granted state-DB access — skip that step if you don't have both.

1. Add the gateway as an MCP server​

claude mcp add --transport http db-gateway --scope project https://db.internal.acme.com

--scope project writes the config into .mcp.json in the current repo so your teammates inherit it on git pull. Use --scope user if it's just for you. Either way nothing sensitive lands in the file — no tokens, no DB info, just the URL.

Verify Claude Code picked it up:

/mcp

You'll see db-gateway listed with status disconnected or needs auth. That's expected — you haven't signed in yet.

2. First call triggers SSO​

Ask the agent anything that needs the gateway:

What databases can I see through db-gateway?

Claude Code calls list_servers, gets back 401 unauthenticated, and surfaces the login link. Click it (or copy-paste into a browser if your terminal can't open URLs). You'll be redirected through your org's IdP — same flow as logging into any internal app.

When the IdP redirects back, Claude Code stores the token in your system keychain (macOS) or credentials file (Linux/Windows) and retries the call automatically.

Expected total time: 10–30 seconds.

3. List what you can reach​

After SSO, the same prompt completes:

You:    What databases can I see through db-gateway?

Claude: [calls list_servers]
        [calls list_databases for each server]

        You have access to:
        • staging (postgres) — staging/app, staging/billing
        • prod (postgres)    — prod/billing (read-only, reason required)

If you see nothing here, the auth worked but you're not in any group that has grants. Skip to Gotchas → I see no servers.

4. Look at a schema​

Pick any database the list showed you. Ask in natural language:

What columns does the users table have in staging/app?

The agent calls describe_schema and renders something like:

staging/app.users

| column     | type        | nullable |
|------------|-------------|----------|
| id         | bigint      | no       |
| email      | text        | no       |
| created_at | timestamptz | no       |
| status     | text        | no       |

Schema lookups are cached by the gateway — re-asking is cheap.

5. Run a query​

How many users signed up in staging/app over the last 7 days?

The agent writes the SQL, picks run_query, and gets back rows + execution stats:

Claude: [calls run_query staging/app]
          select count(*) from users where created_at > now() - interval '7 days';

        → 412 (returned in 38ms)

That's it. You ran a real query against a real DB without ever touching a credential.

6. (Optional) Read your own audit row​

Every run_query call writes a row to the gateway's audit log before the result comes back — user, SQL, reason, row count, duration, outcome.

With the get_query_history tool shipped in #169, ask your agent to call it for the (server, database) you ran against:

get_query_history(server="...", database="...", limit=20)

It returns only your queries (scoped by the SSO-verified identity — no client-supplied user field). If you can't call it, the operator hasn't granted history_read to your group yet; ask the platform team to add one. See ../initial-idea/07-logging-retention.md for the schema and access model.

Gotchas​

I see no servers​

You're authenticated but not a member of any group with a grant. The fix is not in this repo or in Claude Code — it's a PR against the gateway's permissions config. Ping the platform team with: "I logged in as <you@org>, list_servers returns empty. Which group should I be in?"

The browser popup blocker ate the SSO redirect​

Some browsers block window.open() on the first IdP redirect. Two workarounds:

  • Copy the login URL Claude Code prints and paste it into a new tab manually.
  • In Claude Code: /mcp → Reconnect — usually skips the popup and uses your default browser directly.

Corporate proxy / VPN​

If your laptop routes through a TLS-inspecting proxy, the gateway's cert will fail to validate in Claude Code. You'll see a TLS error in the /mcp output, not a 401. Two options:

  • Add your proxy's CA cert to the system trust store (preferred — IT usually has a how-to).
  • Ask your platform team if the gateway is reachable on a path that bypasses the proxy (e.g. a dedicated internal DNS name).

The gateway itself does not support --insecure mode for clients; if cert validation fails it's a real problem to fix, not a flag to flip.

forbidden on a query that used to work​

Two reasons:

  • You were removed from a group. Check with whoever owns the permissions config.
  • The grant was tightened — same group still listed, but the new YAML revoked the action you tried. Permissions are reviewed in git, so git log on the config repo tells you when and why.

reason_required​

Some grants — typically anything against prod — require a one-line reason on every query. Claude Code surfaces this as a prompt:

The gateway requires a reason for this query against prod.
What should I tell it? (e.g. "investigating ticket SUPPORT-4421")

Just type the reason. It lands verbatim in the audit row, so future-you on an audit review will thank present-you for being specific. "checking stuff" is a bad reason; "verifying SUPPORT-4421 root cause" is a good one.

timeout on run_query​

The query exceeded the grant's statement_timeout_ms. Don't try to bypass it — narrow the query (tighter WHERE, smaller table, smaller window) or ask the agent to explain first to see what's expensive.

row_limit_exceeded (or truncated: true in the response)​

The result was clipped at the grant's row_limit. Add a LIMIT, a tighter WHERE, or aggregate (count, sum, group by) before the gateway has to truncate.

"I want to write, not just read"​

By design, every grant defaults to read-only. Writes need (a) an explicit query_write grant on that (server, database) AND (b) a target-DB role with write privileges. Both are operator-side changes, behind a PR review. The gateway will never let an agent self-elevate.

A query_write grant permits data writes only — a single INSERT, UPDATE, or DELETE per call, run through run_query exactly like a read (same statement timeout, row cap, and synchronous audit). Schema changes stay blocked no matter what: CREATE / ALTER / DROP / TRUNCATE, GRANT / REVOKE, and multi-statement bodies are all rejected with forbidden_sql. The gateway never issues DDL.

Token expired silently​

Sessions are short-lived (8h by default). If you see unauthenticated after lunch, that's it — /mcp → Reconnect runs the SSO flow again. Tokens never live in any config file, so there's nothing to clean up.

Gateway tools missing from /mcp entirely​

Either:

  • The gateway URL is wrong — re-add with the right one.
  • The gateway is down — curl -s -o /dev/null -w '%{http_code}\n' https://db.internal.acme.com/readyz should print 200 (checks state-DB reachability, not just liveness).
  • Claude Code hasn't reconnected — restart Claude Code or run /mcp → Reconnect.

Where to go next​