signalpilot

⚡ SignalPilot Data Agent

🏆 #1 on Spider 2.0-DBT — the public state of the art for AI data engineering

65.6 — +5.3 ahead of the next-best system (Databao by JetBrains, May 2026) · 96.9% accuracy on ADE-bench

Governed AI agents for your data stack — db, dbt, and more. Optimized and secured by AutoFyn.

☁️ Try SignalPilot Cloud — free

🚀 Self-host · 📖 Docs · ⭐ Star the repo · 📊 See benchmarks · 🌐 signalpilot.ai · ⚙️ Try AutoFyn · 📅 Book a demo

---

For Agentic Data and Platform Teams

• Governed production access — Snowflake, BigQuery, Postgres, dbt with enterprise guardrails
• Optimized & hardened by AutoFyn — the autonomous security agent that's surfaced vulnerabilities in Next.js, MetaMask, and other major platforms — now tuning prompts, skills, and security guardrails for your data agents
• Benchmark-driven evaluation — same eval rigor that earned #1 on Spider 2.0-DBT
• Enterprise support — SSO, private deployments, SLAs

Talk to us · signalpilot.ai

---

What SignalPilot Is

Two entrypoints, three pieces of infrastructure on the same gateway.

The supported entrypoints are Claude Code and Codex — both via a native plugin. Underneath them, the same components do the work:

1. Plugin (skills + agents) — adds 23 dbt/SQL skills + 2 verifier agents + 40 MCP tools to your agent session. This is the recommended way to use SignalPilot, on Claude Code or Codex.
2. MCP server — standard streamable-http, the layer the plugins talk to. Experimental for other clients: Cursor / custom Agent SDK builds can connect and call the 40 MCP tools, but the skills are plugin-specific and don't run there. Use at your own risk until other platforms ship a skill-equivalent surface.
3. Observability platform — docker compose up -d brings up the gateway, web UI (:3200), audit log, query history, latency/error dashboards, encrypted credential storage, and SignalPilot Workspaces. Or use SignalPilot Cloud for SSO and hosted history.

---

Benchmarks

Most data agents look impressive in a demo and fall apart on a real warehouse. We measure ours against the field — and lead it.

🥇 #1 on Spider 2.0-DBT

The hardest public benchmark for AI data engineering — 68 tasks that demand real project navigation, complex SQL environments, and long-context reasoning. SignalPilot tops the leaderboard at 65.6 — +5.3 ahead of the next-best system (Databao by JetBrains, May 2026), and well clear of the GPT-5-based agents below it. See the leaderboard →

🎯 96.9% on ADE-bench

Across 64 real analytics-engineering tasks — building dbt models against live warehouses — SignalPilot's agents delivered correct, production-ready output 62 of 64 times. Not plausible-looking SQL: models you can ship.

That gap is the difference between an agent you babysit and one you trust in production.

---

Index — What It Is · Benchmarks · How It Works · Try · Architecture · Workspaces · MCP Tools · Community

---

How It Works

Five stages, every task: plan → scan → govern → build → report.

01 — Describe what you need

• 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

02 — Agent scans your project

• 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

03 — Every query is governed

• DDL (DROP, CREATE, ALTER) and DML (INSERT, UPDATE, DELETE) blocked at the parser
• Auto-LIMIT injection on unbounded SELECT
• Per-session budget cap kills queries that would scan over your $ threshold
• Every query audited: timestamp, agent ID, policy reason, full SQL

04 — Plan, build, and verify

• Writes a structured build spec first — sources, joins, grain, filters, expected rows per model — before a line of SQL
• dbt parse catches structural errors; models build in dependency order
• Two read-only subagents verify every built model in parallel — neither edits files, runs dbt, or touches state
• The 7-check verifier audits structure: table existence, column completeness, row count vs source, fan-out, cardinality/grain, non-deterministic SQL, and source-table preservation
• The value-verifier audits values: sample spot-checks, aggregate cross-validation (total vs distinct, measured against source), and status-column filtering for returns/cancellations
• Each verifier returns PASS/FAIL/WARN per check and prescribes the exact CHANGE: fix — no editorializing, numbers measured from source
• The agent applies only FAIL fixes, rebuilds, and re-dispatches both verifiers — looping until every check passes, then stops

05 — Full audit receipt

• Structured summary: duration · agent turns · governed queries · queries blocked · models built · columns validated
• Every line traces back to a specific MCP tool call

---

Try SignalPilot Data Agent

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 Code
claude plugin marketplace add SignalPilot-Labs/signalpilot-plugin
claude plugin install signalpilot-dbt@signalpilot

# (Optional) Install the plugin for skills + agents — Codex
codex plugin marketplace add SignalPilot-Labs/codex-signalpilot-plugin
codex plugin add signalpilot@signalpilot

That's it. Claude Code (or Codex) now has governed access to your databases.

Codex: if you previously added a broken marketplace snapshot, refresh it with codex plugin marketplace upgrade signalpilot && codex plugin add signalpilot@signalpilot.

---

Architecture

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, Codex, Agent SDK, MCP client)  │
└────────────────────────────┬────────────────────────────────┘
                             │ MCP Protocol (streamable-http)
┌────────────────────────────▼────────────────────────────────┐
│  SignalPilot Gateway                                         │
│  ┌────────────┐ ┌──────────────┐ ┌───────────────────────┐ │
│  │ Governance │ │ Schema       │ │ dbt Project           │ │
│  │ • LIMIT    │ │ • DDL        │ │ • Map / Validate      │ │
│  │ • DDL block│ │ • Explore    │ │ • Model verification  │ │
│  │ • Audit    │ │ • Join paths │ │ • Date boundaries     │ │
│  └────────────┘ └──────────────┘ └───────────────────────┘ │
│  ┌────────────────────────┐  ┌────────────────────────────┐ │
│  │ Workspaces (notebooks) │  │ Knowledge base · Notion    │ │
│  │ sandboxed pods (gVisor)│  │ web UI · audit · history   │ │
│  └────────────────────────┘  └────────────────────────────┘ │
└────────────────────────────┬────────────────────────────────┘
                             │  11 governed connectors
   ┌──────────┬──────────┬───┴──────┬──────────┬──────────┐
   ▼          ▼          ▼          ▼          ▼          ▼
 Postgres  Snowflake  BigQuery   DuckDB    Redshift   …+6
 MySQL · SQLite · Databricks · ClickHouse · MSSQL · Trino

---

Plugin (Claude Code)

The SignalPilot plugin adds 23 battle-tested skills + 2 verifier agents to Claude Code and Codex — the same skills that power the Spider 2.0-DBT SOTA.

# Claude Code
claude plugin marketplace add SignalPilot-Labs/signalpilot-plugin
claude plugin install signalpilot-dbt@signalpilot

# Codex
codex plugin marketplace add SignalPilot-Labs/codex-signalpilot-plugin
codex plugin add signalpilot@signalpilot

Core skills: dbt-workflow (full 8-step lifecycle), dbt-write, dbt-debugging, dbt-testing, dbt-snapshots, dbt-versioning, dbt-knowledgebase, sql-workflow, plus dialect skills (duckdb-sql, snowflake-sql, bigquery-sql, sqlite-sql) and domain packs (ecommerce, financial, healthcare, marketing, media, product, hr).

Agents: verifier (post-build 7-check protocol) and value-verifier (aggregate value checks).

---

Use With Any MCP Client

⚠️ Experimental for clients without a plugin. The 40 MCP tools work over streamable-http from any MCP client (Cursor, custom Agent SDK) — but the SignalPilot skills only run inside the Claude Code and Codex plugins. Elsewhere you'll have the tools without skill orchestration. The plugins are the supported path; treat the configs below as best-effort.

Claude Code (one-liner)

claude mcp add --transport http signalpilot http://localhost:3300/mcp

Claude Desktop / Cursor / Any MCP Client

Add 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.

---

Connect a Database

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 (11 connectors): PostgreSQL, MySQL, Snowflake, BigQuery, Redshift, Databricks, DuckDB, SQLite, Microsoft SQL Server, ClickHouse, Trino — plus connection pooling and SSH tunneling.

---

SignalPilot Workspaces

Reactive Python notebooks your agent can drive end-to-end. The run_notebook MCP tool spins up a sandboxed, git-backed workspace; the agent writes and runs cells, and the notebook's Data SDK (import signalpilot as sp) queries your warehouse through the same governed gateway. Open and edit them in the web UI, or let the agent build them headless.

Each workspace runs as an isolated, sandboxed pod (gVisor) with per-org network isolation; projects are git-backed, so no shared filesystem is required.

---

MCP Tools