The Architecture of Trust

1. The Recorder Engine

The heart of EPI is its ability to capture workflow execution into a portable artifact that can be inspected later. Earlier releases introduced explicit capture, and v2.8.7 extends that foundation with stronger agent-first review flows, native OS opening, and deeper framework integrations.

A. Explicit API and Wrappers

EPI works best when recording is explicit. The record(...) context manager, agent_run(...) helpers, and wrapper clients such as wrap_openai() capture inputs, tool activity, outputs, and decisions where you choose to record them. That keeps the product predictable and makes the resulting evidence easier to review.

B. The Security Kernel

Trust is not achieved by logs alone. EPI computes hashes for the sealed files in the artifact and then signs the manifest with Ed25519. The manifest is canonicalized before signing so the same artifact data produces the same verification result across platforms.

1A. New in v2.8.7: Framework Integrations and OS Associations

LiteLLM - 100+ Providers

A single callback can capture LLM calls across 100+ providers: OpenAI, Anthropic, Cohere, Mistral, Azure, Bedrock, and more. This gives teams a practical way to generate evidence without rewriting each provider integration from scratch:

import litellm
from epi_recorder.integrations.litellm import EPICallback

litellm.callbacks = [EPICallback()]  # That's it
response = litellm.completion(model="gpt-4", messages=[...])
# Every call -> signed .epi evidence

LangChain Callback Handler

The EPICallbackHandler captures LLM calls, tool invocations, chain steps, retriever queries, and agent decisions in one handler:

from langchain_openai import ChatOpenAI
from epi_recorder.integrations.langchain import EPICallbackHandler

llm = ChatOpenAI(model="gpt-4", callbacks=[EPICallbackHandler()])
result = llm.invoke("Analyze this contract...")
# Captures: LLM, tools, chains, retrievers, agents

pytest Plugin (--epi flag)

Generate signed .epi evidence per test. This is useful when teams want CI/CD runs to leave behind portable proof instead of only ephemeral logs.

OpenAI Streaming Support

Native stream=True capture lets the caller receive streaming output while EPI records the assembled response and token usage into the artifact.

OpenTelemetry Exporter

EPISpanExporter bridges OpenTelemetry spans into signed .epi files. That makes EPI easier to adopt inside systems that already use OTel for tracing:

from epi_recorder.integrations.opentelemetry import setup_epi_tracing

setup_epi_tracing(service_name="my-agent")
# All OTel spans -> signed .epi files automatically

Global Install

Auto-recording can also be installed globally through sitecustomize.py. One command: epi install --global. Idempotent with clean uninstall via epi uninstall --global.