Framework Integrations
CrewAI, LangChain, OpenAI Agents SDK, HuggingFace, OpenClaw, and Hermes.
Overview
Kalibr integrates with popular agent frameworks and supports any model across any modality. Use router.as_langchain() for LangChain, or auto-instrument HuggingFace's InferenceClient for 17 task types.
CrewAI
Install
shell
pip install kalibr[crewai] crewai openai anthropic
Code (Python only)
python
from kalibr import Router
from crewai import Agent, Task, Crew
# Create Kalibr router
router = Router(
goal="research_task",
paths=["gpt-4o-mini", "claude-sonnet-4-20250514"]
)
# Get LangChain-compatible LLM
llm = router.as_langchain()
# Use with CrewAI
researcher = Agent(
role="Researcher",
goal="Find accurate information",
backstory="You are a research assistant.",
llm=llm, # Kalibr handles model selection
verbose=True
)
task = Task(
description="Research the history of Python programming language.",
expected_output="A summary of Python's history.",
agent=researcher
)
crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()
# Report outcome to Kalibr
router.report(success=len(str(result)) > 100)
Note: CrewAI requires Python 3.10+.
LangChain
Install
shell
pip install kalibr[langchain] langchain langchain-openai langchain-anthropic
Code (Python only)
python
from kalibr import Router
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
router = Router(
goal="summarize_text",
paths=["gpt-4o-mini", "claude-sonnet-4-20250514"]
)
llm = router.as_langchain()
prompt = ChatPromptTemplate.from_template("Summarize this: {text}")
chain = prompt | llm | StrOutputParser()
result = chain.invoke({"text": "Long article text here..."})
router.report(success=len(result) > 50)
OpenAI Agents SDK
Install
shell
pip install kalibr[openai-agents] openai-agents
Code (Python only)
python
from kalibr import Router
router = Router(
goal="agent_task",
paths=["gpt-4o", "gpt-4o-mini"]
)
# Get routing decision
from kalibr import get_policy
policy = get_policy(goal="agent_task")
model = policy["recommended_model"]
# Use with OpenAI Agents SDK
from agents import Agent, Runner
agent = Agent(
name="Assistant",
instructions="You are a helpful assistant.",
model=model,
)
result = Runner.run_sync(agent, "Your task here")
# Report outcome
router.report(success=len(result.final_output) > 0)
HuggingFace
Install
shell
pip install kalibr huggingface_hub
Code (Python only)
python
import kalibr # instruments InferenceClient automatically
from huggingface_hub import InferenceClient
client = InferenceClient()
# All 17 task methods are now traced with cost tracking
# Route between HuggingFace models
from kalibr import Router
router = Router(
goal="transcribe_call",
paths=["openai/whisper-large-v3", "facebook/seamless-m4t-v2-large"],
success_when=lambda output: len(output) > 50
)
result = router.execute(task="automatic_speech_recognition", input_data=audio_bytes)
Note: HuggingFace instrumentation covers 17 task types across text, audio, image, embedding, and classification. Set HF_API_TOKEN or HUGGING_FACE_HUB_TOKEN to access private models or avoid free-tier rate limits.
Supported task types
Pass any of these as the task argument to router.execute():
| Modality | Tasks |
|---|---|
| Text | chat_completion, text_generation, summarization, translation, fill_mask, table_question_answering |
| Audio | automatic_speech_recognition, text_to_speech, audio_classification |
| Image | text_to_image, image_to_text, image_classification, image_segmentation, object_detection |
| Embedding | feature_extraction |
| Classification | text_classification, token_classification |
DeepSeek
DeepSeek uses the OpenAI Python SDK with a different endpoint. No separate SDK needed, just set your API key and use deepseek-* model names in your paths.
Install
shell
pip install kalibr openai
Setup
shell
export DEEPSEEK_API_KEY=sk-... # from platform.deepseek.com
Code (Python only)
python
import kalibr
from kalibr import Router
# Route between DeepSeek models
router = Router(
goal="classify_icp",
paths=["deepseek-chat", "deepseek-reasoner"],
success_when=lambda output: len(output) > 0
)
response = router.completion(messages=[{"role": "user", "content": "Is this company a good ICP fit?"}])
# Mix DeepSeek with other providers, Thompson Sampling learns which wins
router = Router(
goal="classify_icp",
paths=["gpt-4o-mini", "deepseek-chat", "claude-sonnet-4-20250514"],
)
response = router.completion(messages=[...])
Models: deepseek-chat (DeepSeek-V3, fast and cheap), deepseek-reasoner (DeepSeek-R1, for complex reasoning), deepseek-coder (code tasks). Kalibr tracks cost and attributes spans correctly for each.
Voice AI (ElevenLabs, Deepgram, OpenAI Audio)
Install
shell
pip install kalibr[voice] # ElevenLabs + Deepgram + OpenAI Audio pip install kalibr[elevenlabs] # ElevenLabs only pip install kalibr[deepgram] # Deepgram only
TTS routing with router.synthesize()
python
import kalibr # instruments ElevenLabs + OpenAI Audio automatically
from kalibr import Router
# Route between TTS providers, Kalibr learns which sounds better / costs less
router = Router(
goal="narrate",
paths=["tts-1", "eleven_multilingual_v2"],
success_when=lambda _: True # or evaluate audio quality downstream
)
result = router.synthesize("Your agent completed the task.", voice="alloy")
audio_bytes = result.audio
# result.cost_usd, result.model, result.kalibr_trace_id also available
STT routing with router.transcribe()
python
router = Router(
goal="transcribe_calls",
paths=["whisper-1", "nova-2"], # OpenAI Whisper or Deepgram Nova
success_when=lambda text: len(text) > 20
)
result = router.transcribe(audio_bytes, audio_duration_seconds=30.0)
print(result.text) # transcribed text
Auto-instrumentation: ElevenLabs and Deepgram SDKs are patched automatically on import kalibr. All TTS/STT calls are traced with cost tracking (per-character for TTS, per-second for STT).
Voice agent frameworks
python
from kalibr_voice import KalibrLiveKitInstrumentor, KalibrPipecatInstrumentor KalibrLiveKitInstrumentor().instrument() # LiveKit Agents pipeline KalibrPipecatInstrumentor().instrument() # Pipecat processors
Required env vars: OPENAI_API_KEY for Whisper/TTS-1 / ELEVENLABS_API_KEY for ElevenLabs / DEEPGRAM_API_KEY for Deepgram
Voice AI (ElevenLabs, Deepgram, OpenAI Audio)
Route TTS and STT calls across providers using router.synthesize() and router.transcribe(). Same Thompson Sampling outcome loop as text routing.
Install
shell
pip install kalibr[voice] # ElevenLabs + Deepgram pip install kalibr[elevenlabs] # ElevenLabs only pip install kalibr[deepgram] # Deepgram only
TTS routing
python
import kalibr
from kalibr import Router
# Route between ElevenLabs and OpenAI TTS
router = Router(
goal="narrate",
paths=["eleven_multilingual_v2", "tts-1"],
success_when=lambda audio: len(audio) > 0,
)
result = router.synthesize("Hello, welcome to Kalibr.", voice="alloy")
audio_bytes = result.audio
# result.cost_usd, result.model, result.kalibr_trace_id
STT routing
python
import kalibr
from kalibr import Router
# Route between OpenAI Whisper and Deepgram Nova
router = Router(
goal="transcribe_calls",
paths=["whisper-1", "nova-2"],
success_when=lambda text: len(text) > 10,
)
result = router.transcribe(audio_bytes, audio_duration_seconds=30.0)
transcript = result.text
# result.cost_usd, result.kalibr_trace_id
Voice agent frameworks
python
from kalibr_voice import KalibrLiveKitInstrumentor, KalibrPipecatInstrumentor KalibrLiveKitInstrumentor().instrument() # LiveKit Agents pipeline KalibrPipecatInstrumentor().instrument() # Pipecat processors
Install: pip install kalibr[livekit] or pip install kalibr[pipecat]
Provider detection (automatic from model prefix):
- tts-1, tts-1-hd, whisper-1 -- OpenAI Audio
- eleven_* -- ElevenLabs (requires ELEVENLABS_API_KEY)
- nova-*, aura-*, enhanced, base -- Deepgram (requires DEEPGRAM_API_KEY)
When to Use as_langchain() vs get_policy()
| Method | Use When |
|---|---|
| as_langchain() | Framework expects a LangChain LLM (CrewAI, LangChain chains) |
| get_policy() | You need the model name to pass to another SDK |
| router.completion() | Direct text LLM calls without a framework |
| router.execute() | Any HuggingFace task (transcription, image gen, embeddings, etc.) |
Claude Code Plugin
The Kalibr Claude Code plugin instruments your project automatically from within Claude Code.
Install
shell
/plugin add kalibr
Instrument an existing project
shell
/kalibrize
This scans all Python files for bare OpenAI, Anthropic, LangChain, and CrewAI calls, shows each call with a proposed Router replacement, and applies changes with your approval.
What it does automatically
- Injects Kalibr context so Claude Code generates Router-based code when building agents
- Adds kalibr to requirements.txt
- Adds KALIBR_API_KEY and KALIBR_TENANT_ID to .env.example
Next
- API Reference - Full Router API
- Production Guide - Error handling, monitoring