Kestrel - Every failure has a trail.

Four commands. One trail.

Same stack, start to finish. Scroll to watch Kestrel work.

01Inspect

02Watch

03Diagnose

04Verify

Map the stack.

Reads your repo and finds every service, port, health route and dependency. No config, no README archaeology.

kestrel inspect

Run it in order.

Starts services by dependency, probes readiness, classifies logs live and redacts secrets before anything is stored.

kestrel watch

Follow it back.

When things fall over, Kestrel walks the dependency graph and ports to the first failure, and ranks causes with event IDs.

kestrel diagnose

Prove the fix.

A 30 second, multi-sample stability window across every service. Not one lucky health check.

kestrel verify

~/shop-appscanning

webvite · :5173

apiexpress · :4000

workernode · queue

root causedbpostgres · :5432

PASS

$ kestrel inspect ◆ 4 services · vite, express, node, postgres ◆ health GET /health src/server.ts:41 ◆ depends web → api → db ← worker

$ kestrel watch ✓ db ready 1.2s ✓ api ready 2.8s ✓ worker ready 1.9s ✓ web ready 0.9s ◆ redaction on · ring buffer bounded

$ kestrel diagnose 14:02:11.384 db exited 137 ← first signal 14:02:11.512 api, worker ECONNREFUSED :5432 K002 dependency-down confidence 0.94

$ kestrel verify window 30s · 6 samples per service ●●●●●● db api worker web PASS fix holds · 0 orphans

kestrel mcp

Your agent stops guessing.

01/03 You ask your agent what broke. 02/03 It asks Kestrel over MCP. 03/03 Kestrel answers with evidence.

shop-app/ compose.yaml⌘K agent

FILES

api

web

compose.yaml

kestrel.yaml

README.md

server.tscompose.yaml

services:
  db:
    image: postgres:16
    mem_limit: 64m 512m
  api:
    build: ./api
    ports: ["4000:4000"]
    depends_on: [db]
  web:
    build: ./web
    ports: ["5173:5173"]
    depends_on: [api]

Agentkestrel mcp

Why is my app throwing 502s?

⚙ kestrel mcp · 5 tool callsrunning…done in 35ms

db was OOM-killed (exit 137) at 14:02:11. api and web failed after it. I raised mem_limit to 512m and Kestrel verified the fix.

↵

● kestrel mcp connectedread-onlyincident #3main

trace · tools/callsession #3

0102030

get_service_status

4ms

db down · api, web erroring

get_recent_errors

6ms

ECONNREFUSED 127.0.0.1:5432 ×214

get_failure_timeline

9ms

first signal db exited 137 · 14:02:11.384

get_diagnosis

11ms

K002 dependency-down · confidence 0.94

get_verification_results

5ms

PASS · 6/6 samples after the fix

✓ answered with evidence0 raw log lines in context

Logs tell you what happened.Kestrel tells you why.

	Kestrel	Shell / logs	Compose logs	LLM on raw logs
Zero-config execution	Yes	Yes	Needs compose	Yes
Causal root-cause attribution	Graph + ports	No	No	Guesses on timing
Clean process containment	Zero orphans	Leaves orphans	Containers only	No
Multi-sample stability window	Continuous	No	No	No
Automatic secret redaction	Pre-storage	No	No	No
Bounded memory under floods	Ring buffers	Hangs terminal	Fills disk	Token overflow
Offline and private	100% local	Yes	Yes	No

K001

Port conflict

Two things want the same port. Kestrel names both.

K002

Dependency down

A dependency died and took its dependents with it.

K003

Crash after start

Became ready, then fell over.

K004

Startup failure

Never reached ready. Here's the first fatal line.

K005

Dependency not running

Started before what it needs was up.

K006

Missing module

An import or package that isn't installed.

Every failurehas a trail.