Custom Workflows

SignalPilot's 32 MCP tools work in any orchestration — not just the bundled skills. This guide shows how to build your own workflows.

Pick the right tools

Start by understanding what's available:

Task	Tools to use
Orient in a new schema	list_tables, schema_overview, get_relationships
Understand a specific table	describe_table, explore_table, explore_columns
Find how to join two tables	find_join_path, compare_join_types
Write and validate SQL	validate_sql, query_database
Check cost before querying	estimate_query_cost
Debug a dbt error	dbt_error_parser, generate_sql_skeleton
Verify a model build	check_model_schema, validate_model_output, analyze_grain

Structure your prompts

For custom workflows without skills, structure prompts to guide the tool sequence explicitly. Example:

"First use list_database_connections to confirm which connection is active, then schema_overview to understand the database, then find_join_path from orders to products, then write and validate a SQL query using validate_sql before executing."

Or pin this in CLAUDE.md so you don't repeat it every session — see CLAUDE.md recipes.

When to use validators

Always validate before executing on production warehouses:

validation sequence

validate_sql → confirm syntax is clean
estimate_query_cost → confirm cost is acceptable (BigQuery/Snowflake)
query_database → execute with governance

For DuckDB/SQLite/PostgreSQL where cost isn't an issue, you can skip estimate_query_cost and go straight to validate_sql → query_database.

Project-local skills

If you want custom skill behavior for your specific project, add skill files to .claude/skills/ in your project root:

skill directory

your-project/
└── .claude/
    └── skills/
        └── my-custom-skill.md

Claude Code auto-loads files from .claude/skills/ when present. Structure them like SignalPilot's built-in skills: trigger conditions at the top, step-by-step instructions, tool call sequences.

Example .claude/skills/my-custom-skill.md:

custom skill example

---
name: my-etl-workflow
description: "Load when building ETL pipelines for our Snowflake warehouse."
type: skill
---
# ETL Workflow
## When this applies
- User asks to build or update an ETL pipeline
- Snowflake destination is mentioned
## Step 1 — Check source schema
Call `describe_table` on the source table first.
## Step 2 — Validate SQL
Always call `validate_sql` before `query_database`.
## Step 3 — Check cost
Call `estimate_query_cost` before any full-table scan.

Using SignalPilot tools in custom agents

The MCP tools are callable from any Agent SDK that supports streamable-http. If you're building a custom agent:

1. Register the SignalPilot MCP server at http://localhost:3300/mcp
2. Call tools by name: query_database, list_tables, etc.
3. Governance runs server-side — no extra configuration needed

The 32 tools are the same regardless of which client calls them. Only the skill orchestration is Claude Code-specific.

Multi-step query pattern

A robust pattern for ad-hoc analysis:

multi-step query pattern

1. list_database_connections     # confirm which DB is active
2. schema_overview               # understand the shape
3. find_join_path (if joining)   # discover the join path
4. validate_sql                  # check syntax
5. estimate_query_cost           # check cost (Snowflake/BigQuery)
6. query_database                # execute
7. explore_columns (on result)   # profile the output

Wire this into your CLAUDE.md or a project-local skill for consistent behavior across sessions.