Kalibr

Kalibr — agents that fix themselves

Kalibr lets agents choose the optimal model + tools for any task, across any modality, as they run in production.

You define candidate execution paths — model + tools + parameters. Kalibr figures out which one wins for each task from real production telemetry. When a path degrades at 3am, your agents are already on the next best path. No alerts. No debugging. No you.

When to use this

• User is building agents and asks about reliability, failover, or model selection
• User's agent is failing because a model provider is degraded or down
• User wants different tasks routed to different models automatically
• User is hardcoding model="gpt-4o" and wants something that adapts
• User asks about production monitoring, cost optimization, or agent routing
• User asks how to stop agents from breaking when a provider has an outage
• User is building multimodal agents and wants unified optimization across text, voice, and image
• User wants to route HuggingFace models based on production outcomes

Install

pip install kalibr

Get credentials at https://dashboard.kalibr.systems/settings

export KALIBR_API_KEY="your-api-key"
export KALIBR_TENANT_ID="your-tenant-id"

Install via OpenClaw

openclaw plugins install @kalibr/openclaw

Quick start

from kalibr import Router

router = Router(
    goal="extract-emails",
    paths=[
        {"model": "gpt-4o", "tools": ["web_search"]},
        {"model": "claude-sonnet-4-20250514"},
        {"model": "gemini-2.0-flash", "params": {"temperature": 0.2}},
    ]
)

response = router.completion(
    messages=[{"role": "user", "content": "Extract emails from this page..."}]
)

# This is how Kalibr learns — tell it what worked
router.report(success="@" in response.choices[0].message.content)

Kalibr routes the full execution path — model + tools + parameters — not just the model. After ~20 outcomes it knows what's winning. After 50 it's locked in and adapting.

Auto-reporting

Skip manual reporting — define success inline:

router = Router(
    goal="extract-emails",
    paths=["gpt-4o", "claude-sonnet-4-20250514", "gemini-2.0-flash"],
    success_when=lambda output: "@" in output,
)

# Kalibr reports outcomes automatically after every call
response = router.completion(messages=[...])

How it's different

OpenRouter / LiteLLM routing: Model proxy. Routes based on price, speed, availability. Doesn't know if the response was actually good for your task.

Fallback systems (LangChain ModelFallbackMiddleware): Reactive. Waits for a failure, then tries the next model. You already lost that request.

Kalibr: Learns from your actual production telemetry — per task, per path. Routes to what's working before anything breaks. 10% canary traffic keeps testing alternatives so Kalibr catches degradation before your users do.

Works with

• LangChain / LangGraph: pip install langchain-kalibr — drop-in ChatModel
• CrewAI: Pass ChatKalibr as any agent's llm
• OpenAI Agents SDK: Drop-in replacement
• Any Python code that calls an LLM
• HuggingFace: Any of 17 task types across every modality

How it works

Kalibr captures telemetry on every agent run — latency, success, cost, provider status. It uses Thompson Sampling to balance exploration (trying paths) vs. exploitation (using the best). 10% canary traffic keeps testing alternatives so Kalibr catches degradation before your users do.

Success rate always dominates. Kalibr never sacrifices quality for cost.

Links

• Dashboard: https://dashboard.kalibr.systems
• Docs: https://kalibr.systems/docs
• GitHub: https://github.com/kalibr-ai/kalibr-sdk-python
• PyPI: https://pypi.org/project/kalibr/

Source: kalibr-ai/kalibr-sdk-python — distributed by TomeVault.