🏆 #1 on Spider 2.0-DBT — 51.56, +7.45 over JetBrains DataBao (Apr 2026)
Governed AI agents for your data stack — db, dbt, and more. Optimized by AutoFyn.
☁️ Try SignalPilot Cloud — free
🚀 Self-host · 📖 Docs · ⭐ Star the repo · 📊 See benchmarks · 🌐 signalpilot.ai · ⚙️ Try AutoFyn · 📅 Book a demo
- Governed production access — Snowflake, BigQuery, Postgres, dbt with enterprise guardrails
- Harness optimization with AutoFyn — tune prompts, skills, retrieval to your data
- Benchmark-driven evaluation — same eval rigor that earned #1 on Spider 2.0-DBT
- Enterprise support — SSO, private deployments, SLAs
One entrypoint, three pieces of infrastructure on the same gateway.
Today the supported entrypoint is Claude Code. Underneath it, three components do the work:
- Plugin (skill + tool) —
plugin/adds 10 dbt/SQL skills + a verifier agent + 32 MCP tools to your Claude Code session. This is the recommended way to use SignalPilot. - MCP server — standard
streamable-http, the layer the plugin talks to. Experimental for non-Claude clients: Cursor / Codex / custom Agent SDK builds can connect and call the 32 MCP tools, but the skills are Claude Code-specific and don't run there. Use at your own risk until other platforms ship a skill-equivalent surface. - Observability platform —
docker compose up -dbrings up the gateway, web UI (:3200), audit log, query history, latency/error dashboards, encrypted credential storage. Or use SignalPilot Cloud for SSO and hosted history.
Index — What It Is · How It Works · Try · Architecture · MCP Tools · Community
Five stages, every task: plan → scan → govern → build → report.
- Plain-English goal in chat (e.g. "Build
shopify__daily_shop— orders, abandoned checkouts, fulfillment counts by day") - Parsed into a structured task — no SQL written, no warehouse touched yet
- Inspects dbt project + warehouse: sources, staging, marts, missing models
- Flags date hazards (
current_date,now()) - Resolves build order across the DAG — deterministic, not a guess
- DDL (
DROP,CREATE,ALTER) and DML (INSERT,UPDATE,DELETE) blocked at the parser - Auto-
LIMITinjection on unboundedSELECT - Per-session budget cap kills queries that would scan over your $ threshold
- Every query audited: timestamp, agent ID, policy reason, full SQL
dbt parseruns first to catch structural errors- Models materialized in topological order
- Verifier agent reads dbt errors and proposes fixes (renames, missing CTEs, fan-out, date-spine guards)
- Tests run after build and feed back into the loop
- Structured summary: duration · agent turns · governed queries · queries blocked · models built · columns validated
- Every line traces back to a specific MCP tool call
Give your AI agent governed, production-ready access to your data stack — db, dbt, and more. Schema discovery, read-only SQL, dbt project management, all through a single MCP server. No hallucinated tables. No dropped rows. No unbounded queries.
# Start SignalPilot
git clone https://github.com/SignalPilot-Labs/signalpilot.git
cd signalpilot
docker compose up -d
# Web UI available at http://localhost:3200
# Connect the MCP server to Claude Code
claude mcp add --transport http signalpilot http://localhost:3300/mcp
# (Optional) Install the plugin for skills + agents
claude plugin marketplace add SignalPilot-Labs/signalpilot-plugin
claude plugin install signalpilot-dbt@signalpilotThat's it. Claude Code now has governed access to your databases.
Other MCP-DB servers don't enforce LIMIT injection, DDL blocking, dangerous function blocking, or audit logging by default. SignalPilot does — that's why agents on it set the SOTA on Spider 2.0-DBT.
┌─────────────────────────────────────────────────────────────┐
│ Your AI Agent (Claude Code, Agent SDK, any MCP client) │
└────────────────────────────┬────────────────────────────────┘
│ MCP Protocol
┌────────────────────────────▼────────────────────────────────┐
│ SignalPilot Gateway │
│ ┌────────────┐ ┌──────────────┐ ┌───────────────────────┐ │
│ │ Governance │ │ Schema │ │ dbt Project │ │
│ │ • LIMIT │ │ • DDL │ │ • Map / Validate │ │
│ │ • DDL block│ │ • Explore │ │ • Model verification │ │
│ │ • Audit │ │ • Join paths │ │ • Date boundaries │ │
│ └────────────┘ └──────────────┘ └───────────────────────┘ │
└────────────────────────────┬────────────────────────────────┘
│
┌────────────────────┼────────────────────┐
▼ ▼ ▼
┌─────────┐ ┌──────────┐ ┌──────────┐
│ DuckDB │ │ Postgres │ │Snowflake │
└─────────┘ └──────────┘ └──────────┘
The SignalPilot plugin adds battle-tested dbt skills to Claude Code — the same skills that power the Spider 2.0-DBT SOTA.
claude plugin marketplace add SignalPilot-Labs/signalpilot-plugin
claude plugin install signalpilot-dbt@signalpilotIncluded skills: dbt-workflow (full 5-step lifecycle), dbt-write, dbt-debugging, dbt-date-spines, duckdb-sql, snowflake-sql, bigquery-sql, sqlite-sql, sql-workflow
See plugin/README.md for details.
⚠️ Experimental for non-Claude clients. The 32 MCP tools work overstreamable-httpfrom any MCP client (Cursor, Codex, custom Agent SDK) — but the SignalPilot skills are Claude Code-specific and don't run outside it. You'll have the tools without skill orchestration. The Claude Code Plugin is the supported path; treat the configs below as best-effort.
claude mcp add --transport http signalpilot http://localhost:3300/mcpAdd to your MCP config (.mcp.json, .cursor/mcp.json, etc.):
{
"mcpServers": {
"signalpilot": {
"type": "http",
"url": "http://localhost:3300/mcp"
}
}
}When API keys are configured, add an Authorization: Bearer YOUR_API_KEY header.
Via the web UI at http://localhost:3200, or via API:
curl -X POST http://localhost:3300/api/connections \
-H "Content-Type: application/json" \
-d '{
"name": "my-warehouse",
"db_type": "duckdb",
"database": "/path/to/warehouse.duckdb"
}'Supported: DuckDB, PostgreSQL, SQLite, Snowflake, BigQuery.
32 governed tools across query execution, schema discovery, dbt intelligence, and model verification.
| Category | Tools |
|---|---|
| Query | query_database, validate_sql, explain_query, estimate_query_cost, debug_cte_query |
| Schema | list_tables, describe_table, explore_table, explore_column, explore_columns, schema_overview, schema_ddl, schema_link, schema_diff, schema_statistics |
| Relationships | get_relationships, find_join_path, compare_join_types |
| dbt | dbt_error_parser, generate_sql_skeleton, check_model_schema, validate_model_output, audit_model_sources, analyze_grain |
| Operational | list_database_connections, connection_health, connector_capabilities, get_date_boundaries, check_budget, query_history, list_projects, get_project |
See the full tools reference in the docs.
- Read-only governance — queries parsed to AST, DDL/DML blocked, 79+ dangerous functions blocked across 7 dialects
- Tenant isolation — API keys, connections, and audit logs are org-scoped
- Encryption at rest — Fernet (AES-128-CBC + HMAC-SHA256)
- Audit logging — every query logged with PII redaction
- Rate limiting — per-IP, per-key, and per-org with brute-force protection
- Non-root containers — gateway runs as UID 10001
See Security docs for our full security model and vulnerability reporting policy.
SignalPilot/
├── signalpilot/
│ ├── gateway/ # FastAPI backend — MCP server, REST API, governance
│ │ └── gateway/
│ │ ├── api/ # REST API modules
│ │ ├── connectors/ # 11 database connectors + pooling + SSH tunneling
│ │ ├── governance/ # Budget, cache, PII redaction, annotations
│ │ ├── mcp/ # 32 MCP tool definitions (modular package)
│ │ ├── engine/ # SQL validation, LIMIT injection, function denylist
│ │ ├── dbt/ # Project scanning, validation, hazard detection
│ │ ├── db/ # SQLAlchemy ORM models + async engine
│ │ └── auth.py # Clerk JWT (cloud) / local auth + org role enforcement
│ └── web/ # Next.js 16 frontend — 20 pages, Tailwind CSS
│ ├── app/ # App router pages (dashboard, connections, query, etc.)
│ ├── components/ # 20 UI components (sidebar, command palette, etc.)
│ └── lib/ # API client, auth context, hooks
├── plugin/ # Claude Code plugin (10 skills, 1 verifier agent)
│ ├── agents/ # Verifier agent (7-check post-build protocol)
│ └── skills/ # dbt-workflow, sql-workflow, db-specific SQL, etc.
├── sp-sandbox/ # gVisor sandboxed Python execution
├── benchmark/ # Spider 2.0-DBT benchmark suite (SOTA: 51.56%)
└── docker-compose.yml # Full stack: web, gateway, postgres, sandbox
- 🐛 Open an issue — bugs, feature requests, connector requests
- 💬 GitHub Discussions — questions, ideas, show-and-tell
- 🔒 Security policy — report vulnerabilities responsibly
If SignalPilot is useful, please ⭐ — it helps a ton.
Apache 2.0 — see LICENSE.