Say what’s broken the way you’d tell a teammate. Houston investigates across your logs, database, payments and code, then tells you the cause, with every claim tied to the evidence behind it.
@Houston checkout is returning 500s for some customers since about 2pm
Working out which project this is
Investigates across
from the web, Slack or an alert webhook
How it works
Houston takes the first look at every incident, so the person who gets paged starts from an answer instead of a blank terminal.
Type it in Houston, mention it in Slack, or point your monitoring at an alert webhook. Houston works out which project it’s about.
Like a good on-call engineer: it forms hypotheses and checks them against your logs, database, payments, deploys and code. Watch every query live.
A short report: what happened, the likely causes ranked by confidence, and what to do next. Every claim links to what Houston read.
Shows its work
Every sentence in a Houston report carries a citation. Open it and you see exactly what Houston read: the log lines, the query result, the diff. Claims it can’t tie to evidence are cut before you ever see them.
PR #482, merged at 13:52, made shipping_address NOT NULL with no default.34
13:58:07 ERROR 23502null value in column "shipping_address" of relation "orders" violates not-null constraint13:58:07 DETAIL Failing row contains (8841, cus_…, null, …)13:58:31 ERROR 23502 · same constraint · ×37 more
Security
Houston reads production logs, payment data and source code. Every control exists to keep that bounded, visible and in your hands.
Investigation tools only read. Houston refuses keys that grant more than it needs, like a full Stripe secret key.
The model picks a tool; Houston’s runner makes the call. Keys are envelope-encrypted per tenant and decrypted only for the call.
Secrets are redacted, PII can be pseudonymized, and you decide which kinds of data may reach which model provider.
Every tool call, model call and settings change, hash-chained per tenant and exportable as JSONL.
Self-host it, air-gapped if you need to. No telemetry and no calls home. Claude through Anthropic or Amazon Bedrock.
Untrusted text is fenced off in prompts, tool arguments are scope-checked, and reports never load remote images.
Cryptography uses FIPS-approved primitives only: AES-256-GCM, SHA-256, HMAC-SHA-256 and PBKDF2.
Connect your stack in a few minutes. The next incident starts with an answer.
Get started