Quickstart

Connect your agent. Kelet starts finding failure patterns automatically.

Before you start

Sign up at console.kelet.ai, then copy your API key from Settings → API Keys.

Create free account →

Integrate your agent

✦ Use AI (recommended)

The AI coding skill reads your codebase, maps your agent flows, and wires everything — SDK, sessions, signals, and a one-click link to set up evaluators.

Install the skill:

Universal (Claude, Cursor, Copilot, Windsurf…)

npx skills add Kelet-ai/skills

Claude Code

/plugin marketplace add Kelet-ai/skills
/plugin install kelet-integration@kelet-skills

Then tell your agent:

Paste into your AI coding assistant Copy

Integrate Kelet into my app

The skill handles the rest: installs the SDK, maps sessions, instruments LLM calls, proposes signals for your specific failure modes, and generates a deeplink to set up automated evaluators.

Tip

Run the skill in every repo that contains agent code — backend, frontend, and any downstream services.

Auto Instrument

Python

1. Get your API key from the Kelet console under Settings → API Keys.

2. Install the SDK:

pip

pip install kelet

uv

uv add kelet

3. Configure at the top of your agent entry point:

import os
import kelet

kelet.configure(api_key=os.environ["KELET_API_KEY"])

Tip

Use KELET_API_KEY env var — don’t hardcode the key.

4. Wrap sessions — only needed if you own the agent loop (custom chaining, Temporal, raw SDK calls). Skip this if pydantic-ai, LangChain, or LangGraph orchestrates for you.

async def run_agent(session_id: str, user_id: str):
async with kelet.agentic_session(session_id=session_id, user_id=user_id):
result = await agent.run(...)
return result

Note

If you skip agentic_session() when you own the loop, spans appear as unlinked individual traces — no sessions, no RCA.

5. Run your agent. Kelet captures every trace immediately. No further changes needed.

TypeScript

1. Get your API key from the Kelet console under Settings → API Keys.

2. Install the SDK and required OTEL peer dependencies:

npm

npm install kelet \
@opentelemetry/api \
@opentelemetry/instrumentation \
@opentelemetry/exporter-trace-otlp-http \
@opentelemetry/sdk-trace-base \
@opentelemetry/resources

bun

bun add kelet \
@opentelemetry/api \
@opentelemetry/instrumentation \
@opentelemetry/exporter-trace-otlp-http \
@opentelemetry/sdk-trace-base \
@opentelemetry/resources

yarn

yarn add kelet \
@opentelemetry/api \
@opentelemetry/instrumentation \
@opentelemetry/exporter-trace-otlp-http \
@opentelemetry/sdk-trace-base \
@opentelemetry/resources

pnpm

pnpm add kelet \
@opentelemetry/api \
@opentelemetry/instrumentation \
@opentelemetry/exporter-trace-otlp-http \
@opentelemetry/sdk-trace-base \
@opentelemetry/resources

3. Configure at the top of your entry point (before any agent code):

import {configure} from 'kelet';

configure({apiKey: process.env.KELET_API_KEY});

4. Wrap sessions — callback-based (AsyncLocalStorage):

import {agenticSession} from 'kelet';

const result = await agenticSession(
{sessionId: session_id, userId: user_id},
async () => {
return await agent.run(...);
}
);

Note

agenticSession is callback-based, not a context manager. The callback’s scope defines the session boundary.

5. Run your agent. Traces appear in the console within seconds.

What happens next

1

Traces collected

Every LLM call, tool use, and session is captured automatically.

→

2

Signals gathered

User feedback and evaluator scores point Kelet to where things went wrong.

→

3

Findings surface

Recurring failure patterns appear with root causes and suggested fixes.

Open console.kelet.ai after your first run. Traces appear within seconds. Once Kelet sees patterns across sessions, Findings surface with root causes and Prompt Patches.

Next steps

• How Kelet works — understand the analysis pipeline
• Sessions — when and how to use agentic_session()
• Signals — what signals are and why they matter