> ## Documentation Index > Fetch the complete documentation index at: https://schemabrain.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # schemabrain serve > Run the MCP server on stdio against the local store. Exposes the 12 typed MCP tools, writes one audit row per call, optionally enforces the PII policy. # `schemabrain serve` Boots the MCP server. Speaks stdio (the protocol Claude Desktop, Cursor, Windsurf, Claude Code, and the Anthropic SDK demo all use). Reads the local SQLite store; never writes back to the source Postgres. ```bash theme={null} schemabrain serve --url-env DATABASE_URL --store-path ./schemabrain.db ``` Every tool call writes exactly one row to the `mcp_audit` table — see [Audit chain](/mechanism/audit-chain). When `--pii-block` is configured, calls that would touch a blocked category return a `refused` envelope (see [Structured recovery](/mechanism/structured-recovery)). *** ## Source One of these is required: | Flag | Purpose | | ------------------- | ------------------------------------------------------------------------ | | `--url-env VARNAME` | Env var holding the source URL. **Preferred.** | | `--source URL` | Source URL as a named flag. Deprecated when the URL contains a password. | ## Storage | Flag | Default | Purpose | | ------------------- | ------------------ | ------------------------------- | | `--store-path PATH` | `./schemabrain.db` | Path to the local SQLite store. | ## Events | Flag | Default | Purpose | | -------------------- | ----------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------- | | `--events-path PATH` | `$SCHEMABRAIN_EVENTS_PATH` or `~/.schemabrain/events.jsonl` | Path to the JSONL events file the bus appends to. Read with `schemabrain tail`. | | `--no-events` | (off) | Disable event emission entirely. No JSONL file is written; no `server_start` / `server_stop` events. Useful for CI. | ## Audit | Flag | Effect | | ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | `--no-audit` | Disable the `mcp_audit` table writer for this process. Tools still respond; the durable per-call record is suppressed. Use for CI runs or tests where audit writes would clutter a shared store. | ## PII policy | Flag | Default | Purpose | | ----------------- | --------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `--pii-block CSV` | `credential,payment_card,government_id` | Comma-separated `PIICategory` list. When a compiled `get_metric` plan touches one of these, the call returns `refused` with `refusal_reason='pii_blocked'`. Unknown category names abort startup with an error listing the 12 valid values. Pass `''` (empty) to clear the operator policy for this run; the always-on catastrophic-leak floor (`credential` / `payment_card` / `government_id`) still refuses at every read gate and cannot be disabled. PII tags still flow to the audit row. | The default — three catastrophic-leak categories — is the floor every production deployment should keep. Broaden it (`contact,health,financial,...`) for stricter postures. See [PII taxonomy](/mechanism/pii-taxonomy) for the full category list. ## Query safety | Flag | Default | Purpose | | --------------------------- | ------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `--statement-timeout-ms MS` | `30000` (30s) | Postgres-level `statement_timeout` applied to every `get_metric` query. A runaway query aborts with `OperationalError` rather than blocking the process pool. Injected into `connect_args.options` so it **cannot** be overridden via URL query params. Pass `0` to disable (Postgres treats `statement_timeout=0` as unbounded). | | `--max-rows-per-result N` | `10000` | Application-level cap on rows returned by each `get_metric` call. Counts **after** the SQL executes (the source DB still does the full scan) — this is a payload-size guard, not a query-cost guard. Use `--statement-timeout-ms` for the latter. Pass `0` to disable the cap. | *** ## How hosts invoke this You almost never run `serve` directly in a terminal. Instead, your MCP host (Claude Desktop, Cursor, Windsurf, Claude Code) spawns it as a subprocess and speaks stdio across the pipe. The host's config file carries the exact argv. The [Setup](/setup) page covers the per-host config blocks. For verifying the install end-to-end without a host, run `examples/anthropic_demo.py` (see [Manual flow — Anthropic SDK demo](/setup/manual#3-wire-your-own-agent-anthropic-sdk)). *** ## Logging `serve` cannot accept `-v` flags from a host config — there is no command line you control. Use the env-var form instead: ```json theme={null} { "mcpServers": { "schemabrain": { "command": "/usr/local/bin/schemabrain", "args": ["serve", "--url-env", "DATABASE_URL", "--store-path", "/abs/path/to/store.db"], "env": { "DATABASE_URL": "postgresql+psycopg://...", "SCHEMABRAIN_LOG_LEVEL": "INFO" } } } } ``` Valid `SCHEMABRAIN_LOG_LEVEL` values: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL` (case-insensitive). Unknown values fall back to `WARNING` and emit a one-line warning to stderr. Output lands in: * **macOS:** `~/Library/Logs/Claude/mcp-server-schemabrain.log` * **Linux:** the equivalent Claude Desktop log path on your platform. *** ## Examples ### Default policy (recommended) ```bash theme={null} schemabrain serve --url-env DATABASE_URL --store-path ./schemabrain.db ``` The three catastrophic-leak categories block automatically. ### Strict policy ```bash theme={null} schemabrain serve --url-env DATABASE_URL --store-path ./schemabrain.db \ --pii-block credential,payment_card,government_id,contact,health ``` ### Tighter runtime budget than the default Defaults are already safe (30s timeout, 10k row cap). Tighten further for interactive use, where you'd rather see a fast refusal than wait out a slow query: ```bash theme={null} schemabrain serve --url-env DATABASE_URL --store-path ./schemabrain.db \ --statement-timeout-ms 5000 \ --max-rows-per-result 1000 ``` ### Clear the operator policy for a debug session ```bash theme={null} schemabrain serve --url-env DATABASE_URL --store-path ./schemabrain.db \ --pii-block '' ``` This clears any operator-added categories (e.g. `contact`, `health`) so those columns are queryable again. The always-on catastrophic-leak floor (`credential` / `payment_card` / `government_id`) still refuses and cannot be turned off; tags still flow into audit rows regardless. ### CI smoke ```bash theme={null} schemabrain serve --url-env DATABASE_URL --store-path ./schemabrain.db \ --no-audit --no-events ``` *** ## Related The 12 typed tools `serve` exposes. The 12 categories `--pii-block` accepts. Stream events from a running `serve` process. Inspect the `mcp_audit` table `serve` writes to.