Kalibr Documentation

Kalibr is an autonomous routing system that learns which agent execution paths actually succeed in production and routes traffic there in real time.

Why Kalibr?

Most teams hardcode model choices based on benchmarks. Benchmarks don't reflect your data, your prompts, or your definition of success.

Kalibr runs continuous experiments in production:

• Routes traffic across paths you define
• Learns from outcomes you report
• Shifts traffic to what works, away from what doesn't

No manual A/B tests. No spreadsheet tracking. No "we should try Claude for this."

What's a Path?

A path isn't just a model - it's a complete execution configuration:

model + tool + parameters = path

Examples:

• gpt-4o
• gpt-4o + calendar_api
• gpt-4o + calendar_api + {temperature: 0.3}
• claude-sonnet + search_tool + {max_tokens: 1000}

Kalibr learns which full configuration works best for each goal.

from kalibr import Router

router = Router(
    goal="book_meeting",
    paths=[
        {"model": "gpt-4o", "tools": ["calendar_api"]},
        {"model": "gpt-4o", "tools": ["google_calendar"]},
        {"model": "claude-sonnet-4-20250514", "tools": ["calendar_api"]}
    ]
)

response = router.completion(messages=[...])
router.report(success=True)

import { Router } from '@kalibr/sdk';

const router = new Router({
  goal: 'book_meeting',
  paths: [
    { model: 'gpt-4o', tools: ['calendar_api'] },
    { model: 'gpt-4o', tools: ['google_calendar'] },
    { model: 'claude-sonnet-4-20250514', tools: ['calendar_api'] }
  ],
});

const response = await router.completion(messages);
await router.report(true);

Kalibr picks the path, makes the call, and learns from the outcome.