add-semantic-layer

add-semantic-layer

Wire a semantic layer in so it becomes the canonical source of truth for metrics. The agent queries it instead of computing metrics from raw tables.

When to add — and when not to

Only add a semantic layer after nao test shows the agent struggling with metric reliability. Not before.

• Increases reliability and stability (one definition per metric).
• Reduces the scope of answerable questions (anything outside the layer is harder, sometimes impossible).

If failures are concentrated on schema gaps or date logic, a semantic layer doesn't help — fix RULES.md first.

Semantic layer vs metric store: a semantic layer is a file (md/yaml) the agent reads to write its own SQL. A metric store exposes metrics through an API the agent calls (query_metric(...)); the framework converts to SQL. dbt MetricFlow Cloud is a metric store. Snowflake views and nao YAML are semantic layers. Bigger reliability gain with a metric store, bigger scope reduction too.

Step 1 — Pick the tool

Path A — dbt MetricFlow (dbt Cloud with Semantic Layer)

Add to .claude/mcp.json:

{
	"mcpServers": {
		"dbt-mcp": {
			"command": "uvx",
			"args": ["dbt-mcp"],
			"env": {
				"DBT_HOST": "us1.dbt.com",
				"MULTICELL_ACCOUNT_PREFIX": "your_prefix",
				"DBT_TOKEN": "${DBT_TOKEN}",
				"DBT_PROD_ENV_ID": "your_env_id",
				"DISABLE_SEMANTIC_LAYER": "false",
				"DISABLE_DISCOVERY": "true",
				"DISABLE_SQL": "true",
				"DISABLE_ADMIN_API": "true",
				"DISABLE_REMOTE": "false"
			}
		}
	}
}

Substitute MULTICELL_ACCOUNT_PREFIX, DBT_PROD_ENV_ID, DBT_HOST from the user's dbt Cloud account. Set DBT_TOKEN in their shell, not in the file. Restart the session and verify the MCP connects (list_metrics).

dbt Core (local-only) is not supported here — no metric-store API to route through.

Hand off to write-context-rules: in ## Key Metrics Reference, route each MetricFlow metric through query_metric (e.g. MRR → query via dbt MCP query_metric (semantic layer)). In ## Analysis Process, instruct the agent to use semantic-layer tools for known metrics instead of raw tables.

Path B — Snowflake views / semantic views

{
	"mcpServers": {
		"snowflake": {
			"command": "uvx",
			"args": ["mcp-server-snowflake"],
			"env": {
				"SNOWFLAKE_ACCOUNT": "your_account",
				"SNOWFLAKE_USER": "your_user",
				"SNOWFLAKE_PASSWORD": "${SNOWFLAKE_PASSWORD}",
				"SNOWFLAKE_WAREHOUSE": "your_warehouse",
				"SNOWFLAKE_DATABASE": "your_database",
				"SNOWFLAKE_SCHEMA": "your_schema",
				"SNOWFLAKE_ROLE": "your_role"
			}
		}
	}
}

For native semantic views (Cortex Analyst), use the Cortex MCP variant with SEMANTIC_VIEW set. Verify package + env-var names against the latest docs — auth options (key pair / OAuth / password) vary.

Identify the semantic surface (curated views like analytics.metrics.* or native semantic views). Hand off to write-context-rules: in ## Key Metrics Reference, route each metric to its view, never the underlying tables.

Path C — Other (no obvious MCP)

Search the MCP registry, the tool's docs, and the user's installed MCPs. If a fit exists, configure it following the pattern from paths A-B. If not: fall back to Path D (nao semantic files) or build a thin MCP wrapper.

Path D — nao semantic files

For users with no existing semantic layer. One file: semantics/semantic.yaml holding all dimensions and metrics together. Use templates/semantic.yaml.

Walk through dimensions first (slice axes: date, plan, country — capture name, type, description, and allowed values for categoricals), then top metrics (capture name, definition, source table + column + aggregation, grain, dimensions, filters).

Hand off to write-context-rules: in ## Key Metrics Reference, point every metric at semantics/semantic.yaml.

Validate

1. Confirm every metric the user cares about now has a routing rule in RULES.md.
2. nao chat one of their top questions; confirm the agent uses the semantic layer.
3. nao test and compare to the pre-semantic-layer baseline pass rate. Reliability is the only reason to do this — measure it.

Recommend next step

• No tests yet → create-context-tests.
• Reliability dropped → audit-context.
• Otherwise → write-context-rules to refine other sections.

Guardrails

• Only after tests show metric failures. Cite them when the user asks "should we add one?"
• One semantic layer at a time. Two competing layers create MECE violations.
• Don't write RULES.md directly. Hand off to write-context-rules.
• Don't store credentials in .claude/mcp.json. Use ${ENV_VAR}. Add the file to .gitignore if anything sensitive lands there.
• Don't invent metrics for Path D. Only encode what the user defines.

Templates

• templates/semantic.yaml — single-file schema for Path D.