SignalPilot is open-source infrastructure for production-ready data agents: deterministic workflows, verification receipts, continuous evals, a compounding knowledge base, governed warehouse access, and deployment blocks for the tools where teams work.

Does SignalPilot replace my data agent?

No. SignalPilot lets teams keep their existing agent and add maintained workflow, verification, evaluation, knowledge, governance, and deployment blocks underneath it.

How does SignalPilot improve accuracy?

SignalPilot makes exploration, building, verification, and auditing deterministic. It checks schemas and join grain, catches drift with continuous evals, and attaches a verification receipt to the answer.

What are SignalPilot's public benchmark results?

SignalPilot ranks #1 on ADE-Bench at 96.9% (62 of 64 tasks) and #1 on Spider 2.0-DBT at 65.63% (42 of 64 tasks). Public task-level results and transcripts are available on the benchmark pages.

Is SignalPilot open source?

Yes. The building blocks are open source under Apache 2.0, and the control plane can be self-hosted with Docker Compose.

Which agent clients are supported?

SignalPilot provides plugin workflows for Claude Code and Codex, plus governed MCP access that can be used from compatible agent clients.

Open-Source Data-Agent Infrastructure

The building blocks for production-ready data agents.

AI didn’t reduce your work. It moved it: orchestrating harnesses, verifying output. SignalPilot is the infrastructure that makes that easy, and makes the same model more than 2× as correct.

Set up in four steps ↓Book a quick call →

SAME CLAUDE MODEL: 39.5% → 96.9% ACCURATE · #1 ON BOTH PUBLIC BENCHMARKS · EVERY TRANSCRIPT PUBLIC

BLOCK 03 · AUTO EVAL

Ship changes without fear

Every update to your agent stack is evaled before it lands. Nothing breaks quietly.

runscorelearn ⟲

Built by and for people who live in the data stack

A Real Example

Same model. Same prompt. Only one was right.

We asked both for total revenue on a Shopify store’s dbt project.

SHOPIFY · DBT PROJECT

BASE CLAUDE CODE
NO BLOCKS

reads the promptwrites SQLruns itships ✓

REPORTED$2,090,694looked right

WITH SIGNALPILOT
SAME MODEL

explorestests the joinwrites SQLbuildsaudits✗ double-countfixesships ✓

VERIFIED$763,290confidence: high

GOVERNANCE · ROUTED VIA MCP GATEWAYWORKFLOW · DETERMINISTIC, EVERY RUNRECEIPT · ATTACHED TO THE ANSWERKNOWLEDGE BASE · MISTAKE SAVED FOR NEXT RUN

The $1.3M mistake came from base Claude Code, shipped under its own green check. With SignalPilot the model didn’t get smarter. The process refused to let it guess.

See the full example →

The Frustrations

You’re not imagining it.

TRUST

“I can’t trust what it gives back.”

Every answer arrives equally confident, right or wrong. SignalPilot audits before anything ships, and shows you the trail and its confidence on every answer.

MAINTENANCE

“I keep rebuilding my harness.”

Config sprawl, prompt tuning, rewrites with every model release. SignalPilot ships the blocks: maintained and continuously evaluated so you don’t.

The Building Blocks

It’s the age of harness building. These are the blocks.

The hard part isn’t wiring tools. It’s accuracy. Take all six, or start with the one you’re missing.

01 · DETERMINISTIC WORKFLOWS

Explore→Build→Verify

The same discipline every run, shipped as a Claude Code or Codex plugin.

02 · VERIFICATION RECEIPTS

✓✓✓✓✓

attach to your GitHub workflow

Every answer ships with a receipt. Verify your agent’s work at a glance.

03 · AUTO EVAL

run→score→learn⟳

Ship updates to your agent stack without worrying about breaking it. Every change is evaled before it lands.

04 · KNOWLEDGE BASE

Your agent encodes the tribal knowledge: definitions, quirks, decisions. Every run adds to it.

05 · GOVERNANCE

mcp gatewaygoverned accessquery log

Agents reach your warehouse through one MCP gateway. Access governed, every query logged.

06 · DEPLOY WHERE YOUR TEAM LIVES

Slack agentNotion agentyour IDE

The same blocks underneath, wherever work happens.

Every block is open source.

Every Way Your Agent Lies

Every lie has a block that kills it.

Invents columns that don’t exist

every column checked against your schema before the build

Double-counts revenue

join grain measured before writing, audited after

Quietly multiplies rows

output compared against the source table

Uses last quarter’s business terms

today’s values pulled live before filtering

Says “done” without checking

nothing ships until the audit passes

Continuous Evals

A hand-rolled harness is a photograph. SignalPilot keeps re-taking the picture.

Your warehouse keeps moving. A hand-rolled harness breaks randomly and waits for a human to notice and patch it. SignalPilot pairs continuous auto-evals with a compounding knowledge base: drift gets caught the run it appears, every catch becomes a rule, and accuracy only moves one way.

columns get renamed → caught on the next runtables change grain → caught on the next runnew values appear → caught on the next run

The Receipts

Two public benchmarks. #1 on both.

Same Claude model on both runs. The only difference is the blocks. Every transcript is public.

ADE-BENCH · % OF TASKS PASSED

Same model, with and without the blocks.

Claude Code alone

39.5%

+ SignalPilot

96.9%

#1 on the leaderboard · every transcript public →

SPIDER 2.0-DBT · % OF TASKS PASSED

The hardest public data-engineering benchmark.

SignalPilot

65.63%

Databao · JetBrains

60.29%

Shadowfax

41.18%

#1 on the live leaderboard · task-by-task results →

Build vs. Buy

You could build this yourself. Keeping it accurate is the hard part.

Everything we used is published: the workflow, the skills, the verifiers, the eval. Run it against your own harness and see where it breaks. If you’d rather not maintain it, that’s the product.

Run our eval on your harness →Book a quick audit call →

Get Started

Keep your agent. Add the blocks.

We’re data engineers. Claude Code kept handing us confident, wrong numbers on our own pipelines. So we built the blocks that make it stop.

QUICKSTART · UP AND RUNNING IN 4 STEPS

Start the control plane

We host it. Sign up and grab your API key. (Or self-host with docker compose.)

Connect your warehouse

Snowflake, Databricks, Postgres, DuckDB… stored and governed for you.

Install the plugin$ claude plugin marketplace add SignalPilot-Labs/signalpilot-plugin$ claude plugin install signalpilot-dbt@signalpilot

Add the MCP gateway

Every query now routes through governance.

$ claude mcp add --transport http signalpilot https://gateway.signalpilot.ai/mcp --header "X-API-KEY: <your-key>"

Full quickstart →★ Star on GitHub Join the Slack

✓ OPEN SOURCE ✓ APACHE 2.0 ✓ SELF-HOST WITH DOCKER COMPOSE