Logs for the AI era

Stop debugging
by clipboard. Let Claude read the logs.

Kamori is an AI-native log aggregator. Ship your logs to one endpoint — then ask Claude, Cursor, or any MCP client what broke. It queries, correlates, and explains.

Self-host free →Read the docs

Connect in under a minute:npm i @usekamori/sdk

~/kamori · claude-code

youwhat happened to api-gateway around 3pm?

↳ query_logs(service="api-gateway", window="14:50–15:10")

↳ anomaly_hint(metric="error_rate", severity=high)

claude

error rate spiked 783× at 14:58:03.

cause: slow_query: "SELECT * FROM sessions" — full table scan after sessions_idx was dropped.

affected: api-gateway, auth-service. recovery at 15:04.

youIs it everyone or specific users?

↳ search_logs(query="full table scan", service="search")

claude

Only users with >500 saved items. Queries with more than 3 active filters skip the index and do a full scan.

Add a composite index on (user_id, item_count, created_at). Should drop search back under 100 ms.

// instrument once, anywhere

import { installShim } from '@usekamori/sdk'

installShim({ url: 'https://ingest.kamori.ai', token: 'sk_live_…' })

Drop-in SDK

One line to start streaming

Node, Python, Go, Rust, Edge runtimes — same SDK. No agent, no sidecar, no schema design afternoon.

→

Native MCP

Your AI can ask directly

Claude, Cursor, Windsurf — they all speak MCP. Kamori shows up as a first-class tool they can call mid-conversation.

api-gateway · error_rateAnomaly

783× baseline at 14:58:03

Auto-detection

Anomalies in plain English

We flag spikes, regressions, and outlier traces before you open the dashboard — and explain them in sentences, not graphs.

Supported by

Works across leading AI agents and coding editors.

Claude

Claude Code

Cursor

Windsurf

GitHub Copilot

Copilot

OpenAI

Gemini CLI

Gemini

Cline

Roo Code

OpenHands

Codex

Qoder

Trae

Qwen Code

Replit

Kilo Code

OpenCode

Coze

Dify

Flowith

Zencoder

Junie

Manus

Ask directly about your app

Claude Code

already knows your code.
Now give it your logs.

Every service, every error, every trace — in one conversation. Follow an error from your API to your database. Ask why something is slow. Get answers without terminal hopping, without dashboard tabs, without pasting a single line.

Zero friction

One line to start logging.

Run npx kamori and you have an ingest server and an MCP server running locally — no config, no install, no account. Add the SDK to your app and logs flow in seconds.

TypeScript, JavaScript, Python, Go, PHP
Buffered sends with automatic retry
Drop-in console shim — zero code changes
Pino and Winston transports included

SDK docs →

# Start Kamori locally — no install, no account
npx kamori

# Ingest server starts on :3110, MCP server on :3111
# Set a token:
npx kamori --token my-secret

# Or via env var:
INGEST_TOKEN=my-secret npx kamori

Up in five minutes.

Docker, npx, or bare Node — pick your poison.

Send your logs

POST JSON from any service. The SDK handles batching and retries, or send raw HTTP if you'd rather skip the dependency.

Run the server

One container: ingest on port 3110, MCP on port 3111. That's the whole stack — no sidecar, no agent, no config files.

Connect your AI

One command adds Kamori to Claude Code, Claude Desktop, or Cursor. It shows up as a tool set, ready to use.

Ask what broke

"What errors hit the payment service after the last deploy?" Claude chains the right tools automatically and hands you an answer.

1. Send a log

curl -X POST http://localhost:3110/v1/ingest \
  -H "Authorization: Bearer your-secret" \
  -H "content-type: application/json" \
  -d '{"service":"api","level":"error",
       "message":"DB timeout","ms":5012}'

2. Start with Docker

docker run -d \
  -p 3110:3110 -p 3111:3111 \
  -e INGEST_TOKEN=your-secret \
  ghcr.io/usekamori/kamori:latest

3. Connect Claude Code

claude mcp add kamori \
  --transport http \
  http://localhost:3111/mcp

One tool. One job.

Kamori ingests your logs, indexes them, and serves them to your AI. It is not Datadog. It does not try to be.

Fast ingest

Thousands of events per second into SQLite with WAL mode. No external database, no ops burden, no moving parts to break at 3am.

Full-text search

Every log field is indexed. Search by message, service, level, or any field — quoted phrases, AND/OR/NOT operators included.

9 MCP tools

Claude chains them automatically. You ask a question; Kamori handles the tool calls. No prompt engineering required.

Live tail

Stream new events as they arrive over plain HTTP. No WebSocket, no polling loop, no extra process.

SDK for every language

TypeScript, JavaScript, Python, Go, PHP — all with buffering, retry backoff, and a drop-in console shim for zero-config logging.

Auth & rate limiting

Token-based auth per project, per-token rate limits, and HMAC signature verification for Vercel, GitHub, and Render webhooks.

Prometheus metrics

Ingest rate, query latency, and database size at /metrics. Ready to scrape with zero config.

Automatic retention

Set RETENTION_DAYS. Old logs disappear. Export to CSV or NDJSON first if you want to keep them.

Syslog ingestion

Forward from rsyslog or syslog-ng over UDP or TCP. No agent, no sidecar, one config line. OSS / self-hosted only.

Ask once.
Claude does the rest.

Claude doesn't need you to translate your question into a query. Connect Kamori and it picks the right tools, chains them in the right order, and gives you an answer — not a dashboard to click through.

anomaly_hint — Compare recent error rate to 7-day baseline
summarize_errors — Error counts grouped by service and level
watch_logs — Long-poll for new entries — up to 60s
search_logs — Full-text FTS5 search across all fields
tail_logs — Stream from a cursor id — no polling overhead

MCP setup guide →

$ What's the error rate looking like?

anomaly_hint(window_minutes=60)

→ spike_factor: 3.8× above baseline

summarize_errors(since="1h ago")

→ payment: 41 errors / auth: 3 errors

query_logs(service="payment", level="error", limit=5)

→ [5 log entries]

Error rate is 3.8× above normal. All 41 payment errors share message "upstream timeout". Started at 15:23 UTC — 4 minutes after the last deploy.

Claude Skills

One-line workflows for developers.

Install Kamori skills into Claude Code and skip the prompt engineering. Each skill is a pre-built investigation workflow — just describe what you need.

kamori-incident

“Something is wrong with payments”

Chains anomaly detection → error summary → raw logs → distributed trace into a structured incident brief.

kamori-deploy-check

“Did the deploy break anything?”

Compares error rates before and after a deploy timestamp. Returns a clear PASS / WARN / FAIL verdict.

kamori-debug-error

“I'm seeing: connection refused”

Paste any error message. Claude finds it in logs, determines frequency, and traces it back through your services.

kamori-monitor

“Watch auth for errors”

Live log tail with anomaly alerting. Calls out error bursts and new patterns as they arrive.

kamori-performance

“Why is the API slow?”

Finds slow requests, computes P50–P99 latency, and traces the slowest request to find the bottleneck service.

kamori-digest

“Weekly log summary”

Generates a daily or weekly health narrative: error trends, anomalies, top patterns, volume overview.

$cp -r kamori-skills/* ~/.claude/skills/

Open source · install in 30 seconds · view all skills →

Stop copy-pasting logs into Claude.

Pipe them in directly. Ask what broke. Kamori handles the rest.

Self-host free

Stop debuggingby clipboard. Let Claude read the logs.