# SignalPilot

> SignalPilot is the self-improving Context Factory for your business data. We ingest your dbt project, BI tools, dashboards, and agent traces to draft your context, metrics, and evals as code, then keep your dbt models, semantic models, and skills correct as your teams use them, with evidence-backed answers from Claude Cowork, Codex, Slack, or the web.

## What is SignalPilot

SignalPilot is the self-improving Context Factory for a company's business data. SignalPilot helps a team curate a context layer that is good enough to start, by ingesting its BI tools, dashboards, dbt project, and agent traces, and then keeps the whole semantic layer correct over time: dbt models, semantic models, and skills, all checked by evals.

The four stages:
- **Ingest and draft** — SignalPilot ingests the dbt project, BI tools, dashboards, and agent traces and drafts the context, metrics, and evals as code; the data team curates the draft
- **Check every change** — evals rerun on every pull request, with blast radius and gating
- **Automatically surface issues in chat** — agents parse chat traces, cluster the context inconsistencies they surface, teammates propose fixes and bump them up, and data context admins review and approve the context changes
- **Act anywhere** — one shared, sandboxed cloud agent on the warehouse's own permissions, reachable from Claude Cowork, Claude Code, Codex, Slack, or the web; evidence graph on every answer
- **Managed** — a SignalPilot forward deployed engineer keeps the Context Factory correct every week, with a 90% precision guarantee

How a deployment runs: SignalPilot ingests the whole estate; the team curates the first draft; evals gate every change; teams settle terms in chat; the sandboxed cloud agent rolls out to one team first. On every data-model change the affected evals rerun; corrections become proposed updates and regression tests, reviewed before they join the system.

Open standards: built on dbt; metric definitions in the Apache Ossie format; business context in plain Markdown, in the customer's repo. Open source under Apache 2.0; self-hostable with Docker Compose.

## Key Features

1. Business context, metrics, and evals drafted as code from your dbt project, BI tools, dashboards, and agent traces
2. Team Context Curation: agents surface ambiguous terms in chat, teammates answer and vote, and the data team merges named, owned definitions
3. Evals on every data-model change with pull request gating and blast radius
4. Evidence graph behind every answer: query, source tables, transformations, assumptions, checks
5. Sandboxed cloud agent reachable from Claude Cowork, Codex, Slack, or the web through an MCP app
6. MCP gateway with a full audit log, on your warehouse's permissions
7. Managed plan with a 90% precision guarantee
8. Open standards: dbt, Apache Ossie metric definitions, Markdown context
9. Open-source components with self-hosting through Docker Compose
10. #1 on ADE-Bench at 96.9% and Spider 2.0-DBT at 75%

## Use cases

- [Team Context Curation](https://signalpilot.ai/solutions/self-improving-context) — Every chat improves your context.
- [Sandboxed Agent Harness](https://signalpilot.ai/solutions/sandboxed-cloud-agent) — One multiplayer agent for everyone.
- [Model & Vendor Bakeoff](https://signalpilot.ai/solutions/bakeoff) — Judge models and save on token cost.

## Who it is for

- [For Data Leads](https://signalpilot.ai/for/data-leads) — Be the orchestrator, not the bottleneck.
- [For Sales, Product & Execs](https://signalpilot.ai/for/business-teams) — Confident actions, from any chat.
- [For Data Consultants](https://signalpilot.ai/for/consultancies) — Build AI projects and recurring context services.
- [For data teams](https://signalpilot.ai/for/data-teams): Accurate, up-to-date context for your data team and AI agents.

## Benchmarks

- Spider 2.0-DBT: #1. 75% (48 of 64) submitted in August 2026; the official leaderboard shows 65.6% (May 2026). Next best: Databao by JetBrains 60.29%, Shadowfax 41.18%, Spider-Agent-Extended with GPT-5 39.71%.
- ADE-Bench (dbt Labs): 96.9% (62 of 64), the highest score recorded. Snowflake Cortex Code CLI 65.1%, dbt Labs 58.1%, Claude Code alone 39.5%. Both SignalPilot and Claude Code runs used Claude Code with Sonnet 4.6.
- Every transcript is public at https://signalpilot.ai/benchmark and https://signalpilot.ai/benchmark/ade.

## Pricing

- Team: $100 a month, includes 5 seats, then $15 per seat; $50 monthly usage credit; 30 data models with evals included
- Scale: $250 a month, includes 15 seats, then $15 per seat; $100 monthly usage credit; 50 data models with evals included; one engineer-week of integration on annual
- Enterprise: custom price per account on one purchase order, includes 100 seats, more by agreement; $1,000 monthly usage credit; 100 data models with evals included; SSO and SCIM, PrivateLink, customer-managed keys, audit export, 99.9% uptime SLA
- Usage on every plan: $0.50 per successful thread plus $0.01 per governed query (failed threads are never billed); $6 per data model with evals per month after the allowance; 30 scheduled eval runs included, then $0.50 plus queries per run; model tokens passed through at the provider's list price with no markup
- Managed: from $2,000 (Team), $4,000 (Scale), $8,000 (Enterprise) a month; a SignalPilot forward deployed engineer runs the weekly loop with a 90% precision guarantee on graded questions; refunds cover SignalPilot fees, never model tokens

Estimate a bill and payback: https://signalpilot.ai/pricing/calculator

## Onboarding

The integration week: a SignalPilot engineer or certified partner works on the customer's schema under NDA. Day 1 scope, day 2 ingest everything, day 3 curate the draft, day 4 evals and gate, day 5 ship. Included on annual Scale and Enterprise plans; quoted per engineer-week on Team, or delivered by a partner. Self-serve quickstart:

1. **Start the control plane**: Sign in to the hosted SignalPilot control plane and get an API key, or self-host with Docker Compose.
2. **Connect your warehouse**: Connect Snowflake, Databricks, Postgres, DuckDB, or another supported warehouse.
3. **Install the plugin**: claude plugin marketplace add SignalPilot-Labs/signalpilot-plugin then claude plugin install signalpilot-dbt@signalpilot
4. **Add the MCP gateway**: claude mcp add --transport http signalpilot https://gateway.signalpilot.ai/mcp --header "X-API-KEY: <your-key>"

## Frequently Asked Questions

### Is this another agent?

No. The sandboxed cloud agent is only a part of the whole offering. SignalPilot is primarily a development lifecycle management platform for your data context and agents: a Context Factory that keeps your context, metrics, and evals correct, plus a cloud sandbox for the agents that use them. Once set up, you can use these sandboxed agents from Claude Cowork, Codex, or Slack.

### Why not just use Claude Code?

Claude Code makes one person faster. Data work at a company is mostly alignment of controlled business vocabulary and procedures. SignalPilot is multiplayer. Teams settle definitions in chat, your data team merges them, and every person and agent answers from the same agreed context.

### Do we have to document everything first?

No. Manually writing definitions, semantic layer, and skill files doesn't scale, so we ingest your dbt models, BI tools, dashboards, and agent traces and draft the context for you. Your team curates the draft, and we help you keep it accurate as your business changes.

### Which warehouses?

Snowflake, Databricks, BigQuery, Postgres, DuckDB, Redshift, and more.

### Where does our data go?

Queries run in a SignalPilot cloud sandbox against your warehouse, on your warehouse's own permissions. Definitions live in dbt, Apache Ossie, and Markdown in your repo. Self-host with Docker Compose if you prefer.

### What does it cost?

From $100 a month for a five-seat team, then $0.50 per successful thread and $0.01 per governed query.

## Supported Data Platforms

- Snowflake
- Databricks
- BigQuery
- Postgres
- DuckDB
- Redshift
- dbt Core and dbt Cloud

## Harnesses and surfaces

- Claude Cowork
- Claude Code
- Codex
- Slack
- Web
- MCP connections provisioned centrally: Salesforce, PostHog, Segment, Notion, Slack

## Team

- Tarik Moon, CEO. Led data science at EDO; data at Goldman Sachs and eBay; founded Affine. Harvard, Math and CS. IMO medalist.
- Fahim Aziz, Product and GTM. Founded Backpack (Y Combinator S14); led growth at Nala.
- Adib Hasan, Engineering. MIT, BA and MEng in Math and CS; previously Meta and Kensho; LLM research at MIT. IMO medalist.
- Daniel Schaffield, Engineering. Senior engineer at Affine.

## Site Pages

- [Homepage](https://signalpilot.ai)
- [Overview](https://signalpilot.ai/product) — The context and the agent, as one loop.
- [Context & Evals](https://signalpilot.ai/product/context-and-evals) — Keep models, metrics, and skills accurate.
- [Sandboxed Cloud Agent](https://signalpilot.ai/product/cloud-agent-harness) — One multiplayer agent, in any chat.
- [Managed](https://signalpilot.ai/product/managed) — Our engineers run your Context Factory.
- [Onboarding](https://signalpilot.ai/product/onboarding) — Your Context Factory, live in a week.
- [Pricing](https://signalpilot.ai/pricing) — Team, Scale, Enterprise; $0.50 per successful thread, $0.01 per governed query
- [Pricing calculator](https://signalpilot.ai/pricing/calculator) — bill estimate and engineering hours saved, all assumptions editable
- [Team Context Curation](https://signalpilot.ai/solutions/self-improving-context) — Every chat improves your context.
- [Sandboxed Agent Harness](https://signalpilot.ai/solutions/sandboxed-cloud-agent) — One multiplayer agent for everyone.
- [Model & Vendor Bakeoff](https://signalpilot.ai/solutions/bakeoff) — Judge models and save on token cost.
- [For Data Leads](https://signalpilot.ai/for/data-leads) — Be the orchestrator, not the bottleneck.
- [For Sales, Product & Execs](https://signalpilot.ai/for/business-teams) — Confident actions, from any chat.
- [For Data Consultants](https://signalpilot.ai/for/consultancies) — Build AI projects and recurring context services.
- [Benchmarks](https://signalpilot.ai/benchmark) — #1 Spider 2.0-DBT · 96.9% ADE-Bench.
- [Case archive](https://signalpilot.ai/comparison) — 39 tasks Claude Code missed, side by side.
- [How we beat JetBrains to #1](https://signalpilot.ai/blog/how-we-beat-jetbrains-to-1-on-the-worlds-hardest-data-benchmark) — Spider 2.0-DBT, the hardest data benchmark.
- [ADE-Bench report: 96.9%](https://signalpilot.ai/blog/signalpilot-ade-bench-96-percent) — The highest score ever, task by task.
- [Build vs. buy](https://signalpilot.ai/compare) — hand-built harnesses versus SignalPilot
- [Live onboarding session](https://signalpilot.ai/akasa-demo) — an unedited data-agent session
- [Quickstart](https://signalpilot.ai/quickstart) — self-serve setup on the open-source stack
- [Blog](https://signalpilot.ai/blog)
- [About](https://signalpilot.ai/about)
- [Partners](https://signalpilot.ai/partners)
- [Support](https://signalpilot.ai/support)
- [Privacy Policy](https://signalpilot.ai/privacy)
- [Terms of Use](https://signalpilot.ai/terms)

## Blog Posts

- [Our Learnings from Beating Both Spider 2.0 and ADE Bench: Plausibly Wrong Is Worse Than No Agent at All](https://signalpilot.ai/blog/correct-not-plausible) — 2026-05-29 — In production, an AI data agent that's plausibly wrong is worse than no agent at all: inflated revenue, dropped customers, double-counted metrics that pass review and surface in a board deck weeks later. SignalPilot produces data that's correct, not just plausible, and stays safe on a real warehouse because AutoFyn, the security agent behind disclosed Next.js and MetaMask bugs, audits it without stopping. The benchmarks (#1 on Spider 2.0-DBT, 96.9% on ADE-Bench) just prove it.
- [SignalPilot ADE-Bench Report: 96.9%, Highest Score Ever](https://signalpilot.ai/blog/signalpilot-ade-bench-96-percent) — 2026-05-28 — SignalPilot resolves 62 of 64 ADE-Bench tasks, the highest score on dbt Labs' analytics engineering benchmark. Three skills and one sentence took us from 71% to 96.9%.
- [How We Beat JetBrains to #1 on the World's Hardest Data Benchmark](https://signalpilot.ai/blog/how-we-beat-jetbrains-to-1-on-the-worlds-hardest-data-benchmark) — 2026-04-24 — Today, we're thrilled to announce that SignalPilot has claimed the #1 spot on the Spider 2.0-DBT leaderboard — beating JetBrains' Databao by over 7 points. By solving dbt, the hardest problem in the transformation layer, we proved AI agents can be trusted with enterprise data pipelines when they have the right guardrails.
- [How Claude Code + SignalPilot Builds Production-Grade dbt Models](https://signalpilot.ai/blog/claude-code-signalpilot-production-dbt) — 2026-04-28 — From empty directory to 16 models, 25 passing tests, and full documentation — in a single conversation. A step-by-step trace of how Claude Code and SignalPilot MCP built a complete dbt project.
- [Benefits of Using SignalPilot Governed MCP to Build dbt Models](https://signalpilot.ai/blog/governed-mcp-for-dbt-models) — 2026-04-28 — Why governed database access through MCP changes the AI-assisted analytics engineering workflow — from AST-level query parsing to automatic LIMIT injection and full audit trails.

## Contact

- Email: tarik@signalpilot.ai
- GitHub: https://github.com/SignalPilot-Labs/SignalPilot
- Slack Community: https://join.slack.com/t/signalpilotcommunity/shared_invite/zt-3w68aak3n-c6_Mks32LtgkAyWxUVp2cw
- X: https://x.com/signalpilot_ai
- LinkedIn: https://www.linkedin.com/company/signalpilot/