Auto-instrumentation

Drop-in integrations for 7 agent frameworks. One import, zero manual wrapping.

Why auto-instrumentation?

Manually wrapping every LLM call is fragile — you miss calls inside library internals, nested chains, and background threads. Framework hooks sit at a lower level: they fire on every model call, tool execution, and chain step whether you wrote that code or not. The result is 100% coverage with a single setup line, and no diff noise when you add new tools or model calls later.

Supported frameworks

Framework	Language	Install	Status
LangChain	TypeScript + Python	npm install @cisora/sdk \| pip install cisora[langchain]	Shipped
LlamaIndex	TypeScript + Python	npm install @cisora/sdk \| pip install cisora[llamaindex]	Coming soon
AutoGen	Python	pip install cisora[autogen]	Coming soon
CrewAI	Python	pip install cisora[crewai]	Coming soon
OpenAI Agents SDK	TypeScript + Python	npm install @cisora/sdk \| pip install cisora[openai-agents]	Coming soon
Mastra	TypeScript	npm install @cisora/sdk	Coming soon
Pydantic AI	Python	pip install cisora[pydantic-ai]	Coming soon

LangChain is fully documented at /docs/frameworks/langchain. The sections below preview the integration pattern for each upcoming framework.

LlamaIndex

Coming soon

Instrument every query engine, retriever, and LLM call across LlamaIndex TS and Python.

Install

npm

npm install @cisora/sdk

pip

pip install cisora[llamaindex]

Usage

TypeScript

import { Settings } from 'llamaindex';
import { CisoraInstrumentation } from '@cisora/sdk/llamaindex';

CisoraInstrumentation.instrument(Settings, {
  apiKey: process.env.CISORA_API_KEY!,
  agentName: 'rag-agent',
});

// All subsequent queries are captured automatically.

Python

from llama_index.core import Settings
from cisora.llamaindex import CisoraInstrumentation

CisoraInstrumentation.instrument(
    api_key=os.environ["CISORA_API_KEY"],
    agent_name="rag-agent",
)

# All subsequent queries are captured automatically.

What gets captured

Every LLM call (query, chat, complete) as a model_call event
Retriever queries and retrieved node counts
Query pipeline steps nested as parent/child spans
Token usage extracted from provider responses

AutoGen

Coming soon

Capture every message exchange and tool call in Microsoft AutoGen multi-agent conversations.

Install

pip

pip install cisora[autogen]

Usage

Python

import autogen
from cisora.autogen import CisoraListener

listener = CisoraListener(
    api_key=os.environ["CISORA_API_KEY"],
    agent_name="multi-agent-team",
)

assistant = autogen.AssistantAgent(
    name="assistant",
    llm_config={"model": "gpt-4o"},
    event_listeners=[listener],
)

What gets captured

Every agent-to-agent message as a conversation_turn event
Tool calls and results surfaced as tool_action events
GroupChat orchestration steps with speaker attribution
Per-agent token usage and cost tracking

CrewAI

Coming soon

Trace every task, agent handoff, and tool execution inside a CrewAI crew.

Install

pip

pip install cisora[crewai]

Usage

Python

from crewai import Crew
from cisora.crewai import instrument_crew

crew = Crew(agents=[...], tasks=[...])
instrument_crew(
    crew,
    api_key=os.environ["CISORA_API_KEY"],
    agent_name="content-crew",
)

result = crew.kickoff()

What gets captured

Every task start and completion as a task_lifecycle event
Agent-to-agent delegation as handoff events
Tool calls with input/output payloads
Crew-level token usage rolled up across all agents

OpenAI Agents SDK

Coming soon

One processor attaches to the OpenAI Agents SDK tracing pipeline — no per-agent setup.

Install

npm

npm install @cisora/sdk

pip

pip install cisora[openai-agents]

Usage

TypeScript

import { addTraceProcessor } from '@openai/agents';
import { CisoraTraceProcessor } from '@cisora/sdk/openai-agents';

addTraceProcessor(
  new CisoraTraceProcessor({
    apiKey: process.env.CISORA_API_KEY!,
    agentName: 'customer-support',
  })
);

// All agents in this process are now instrumented.

Python

from agents import add_trace_processor
from cisora.openai_agents import CisoraTraceProcessor

add_trace_processor(
    CisoraTraceProcessor(
        api_key=os.environ["CISORA_API_KEY"],
        agent_name="customer-support",
    )
)

What gets captured

Every model call including response streaming
Handoffs between agents with source and target labels
Tool calls (function tools, hosted tools, computer-use)
Guard rail results and any tripwire events

Mastra

Coming soon

Install

npm

npm install @cisora/sdk

Usage

TypeScript

import { Mastra } from '@mastra/core';
import { CisoraExporter } from '@cisora/sdk/mastra';

const mastra = new Mastra({
  telemetry: {
    exporter: new CisoraExporter({
      apiKey: process.env.CISORA_API_KEY!,
    }),
  },
  agents: { myAgent },
});

What gets captured

Agent runs, step spans, and LLM calls as nested trace spans
Workflow execution with per-step latency and status
Tool and integration calls with input/output payloads
Memory read/write operations surfaced as retrieval events

Pydantic AI

Coming soon

Attach a Cisora instrument to any Pydantic AI agent — structured outputs and tool calls all captured.

Install

pip

pip install cisora[pydantic-ai]

Usage

Python

from pydantic_ai import Agent
from cisora.pydantic_ai import instrument_agent

agent = Agent('anthropic:claude-sonnet-4-5', result_type=MyResult)
instrument_agent(
    agent,
    api_key=os.environ["CISORA_API_KEY"],
    agent_name="structured-extractor",
)

result = await agent.run("Extract data from this doc …")

What gets captured

Every model request and structured response as a model_call event
Tool calls with typed input schemas and validated output
Retry attempts and validation errors as separate span events
Result type and validation status logged alongside token usage

← SDK reference LangChain guide →