mirrorkit 0.1.1

Mirrors: build a high-fidelity simulation of a production environment from traces (and optional LangGraph code) to run and test agents with no external dependencies.

Mirrors

Build a runnable simulation of a customer's production agent environment from their traces, then browse its assets, run evals, and play with it — all in a web app. This monorepo holds four independently-hostable parts:

┌─────────────┐   collect (Bearer key)   ┌──────────────────────────┐
│  collector/ │ ───────────────────────► │        backend/          │
│ (customer's │                          │  control-plane API       │ ──► Postgres
│   machine)  │                          │  (routing, orchestration)│ ──► ECS Fargate envs
└─────────────┘                          └───────────▲──────────────┘     (built on mirrorkit/)
                                                      │ REST (one origin)
                                          ┌───────────┴──────────────┐
                                          │        frontend/         │  (Vercel)
                                          └──────────────────────────┘

Part	Dir	What it is	Hosted on
Frontend	frontend/	React + Vite app — Dashboard, Ingest, Simulation, Evals, Playground	Vercel
Backend	backend/	FastAPI control plane: routing, env registry, Fargate orchestration, evals/playground, setup assistant, collector ingest. Depends on core mirrorkit.	EC2 + external Postgres
Core engine	mirrorkit/	The simulation engine the backend runs: traces → schema + seed + tool bindings → a runnable Environment; serve + package it. See mirrorkit/README.md.	inside each env container
Collector	collector/	A 2-line, zero-dependency SDK that streams an agent's production traces to the backend. See collector/README.md.	customer's machine

The collector (2 lines)

import mirror
mirror.init(api_key="mk_live_...", project="my-agent")

Auto-instruments LangChain/LangGraph, Anthropic, and OpenAI if present; captures each run's messages + tool calls and ships them in the background (negligible overhead, never raises into your app).

Run it locally

# backend (SQLite for local; or point MIRROR_DATABASE_URL at any Postgres)
MIRROR_DATABASE_URL=sqlite:///data/mirror.db \
  uvicorn backend.app:create_app --factory --port 8001

python examples/seed_demo.py              # register the airline demo env
cd frontend && npm install && npm run dev # http://localhost:5173

The airline demo (agent + traces + seed) lives in examples/. For real (non-mock) data, set ANTHROPIC_API_KEY + MIRROR_SIM_MODEL + MIRROR_BUILD_MODEL (see backend/.env.example).

Deploy

DEPLOY.md — backend on EC2 + an external Postgres, ECS Fargate for the per-customer env containers, Vercel for the frontend. SWITCH.md is the architecture deep-dive.

Develop / test

pip install -e ".[backend,langgraph,serve,backends,dev]"   # core + backend
pip install -e collector                                    # the SDK
pytest -q                  # core + backend tests
pytest collector/tests -q  # collector tests

Install profiles

• Core engine only: pip install -e . (extras: serve, langgraph, backends).
• Backend (adds the control plane): pip install -e ".[backend,langgraph]".
• Collector: pip install -e collector (zero required deps).