Stop asking the agent
how it thinks it’s doing.
Signal House is a local operator dashboard for AI-assisted development. It reads Git, GitHub and agent sessions, then shows the boring evidence that actually matters: delivery, cycle time, CI health and spend.
Busy is not the same
as shipping.
Agent tooling is excellent at producing activity. Sessions run. Tokens disappear. Terminals scroll. None of that tells you whether useful work is getting merged.
Signal House ignores the self-report and reads the systems of record instead. Git tells it what changed. GitHub tells it what shipped. CI tells it what broke. Agent databases tell it what the work cost.
git + githubDid work actually land?observedissues + PRsHow long does delivery take?measuredactions + checksAre changes staying healthy?ground truthsession DBsWhat did that output cost?direct / estimatedany sourceDo we actually know?— unknownRead the ground truth.
Keep it local.
The collector layer reads local repositories, GitHub and supported agent databases. A small orchestration layer refreshes and normalises that data into SQLite. The React dashboard reads one privacy-filtered API surface.
collect → normalise → filterlocal onlyReact + EChartsBuilt for operators,
not dashboard theatre.
The product is intentionally opinionated about data quality and privacy. A prettier number is never more valuable than an honest one.
Unknown stays unknown.
Missing data is rendered as an em dash instead of silently becoming zero. No synthetic zero-days. No confident fiction.
Privacy fails closed.
If repository visibility cannot be verified, Signal House treats it as private. Private-repo attention items remain hidden unless explicitly enabled.
Local means local.
No hosted analytics service, account or telemetry pipeline. One process reads the state already on your machine and serves the operator UI.
Five numbers.
Drill down when needed.
The top-level view is deliberately compact: throughput, cycle time, CI pass rate, stale work and agent spend. The rest of the interface exists to answer “why?” when one of those moves.

Delivery
Commits, merged PRs and CI pass rate share the same 7/30/90-day window so activity and quality can be read together.
Agent spend
Cost, sessions, tokens, cache hit rate and per-model efficiency expose where the budget is actually going.
Attention queue
Open issues and PRs that need eyes surface with age, staleness and CI state attached.
Diagnostics
Collector health, discovered repositories and configuration state stay inspectable instead of failing invisibly.
One runtime.
Very little ceremony.
Signal House was rewritten around Bun to keep deployment boring: one runtime, one package manager and one server process.
git clone git@github.com:barkley-assistant/signal-house.gitcd signal-house && bun installcp .env.example .envbun run devhttp://localhost:3000See whether the work
is actually moving.
Run it locally, inspect the source, or open the live deployment.
