Quickstart: test an AI agent

By the end of this page your agent’s real traces are streaming to Mirrors, you have a runnable environment built from them, and an eval set that scores every future change against recorded sessions.

How do I set up Mirrors, step by step?

Five steps: connect an MCP client so your AI does the driving, add the collector so real sessions start streaming, then build an environment, author an eval set, and read your first result. Steps 3 to 5 can be run from your MCP client, the CLI, or the dashboard; they all hit the same backend.

Install the MCP

Mirrors hosts an MCP server that exposes the full product surface (build, explore, query, and eval environments), so any MCP client can drive it. Point your client at the hosted endpoint; sign-in happens in the browser on first use, no API key to paste.

https://api.runmirrors.com/mcp

claude mcp add --transport http mirrors https://api.runmirrors.com/mcp

Then run /mcp → mirrors → Authenticate via browser.

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "mirrors": {
      "url": "https://api.runmirrors.com/mcp"
    }
  }
}

Cursor shows Needs login on the server. Click it to sign in via the browser.

code --add-mcp '{"name":"mirrors","type":"http","url":"https://api.runmirrors.com/mcp"}'

VS Code prompts to trust the server, then opens the browser to sign in.

On chatgpt.com:

Settings → Apps & Connectors → Advanced → enable Developer mode
Create → name it Mirrors, MCP server URL https://api.runmirrors.com/mcp
Authentication: OAuth → Create, then connect

Sign-in opens in the browser. Needs a Plus/Pro/Business plan (web).

codex mcp add mirrors --url https://api.runmirrors.com/mcp
codex mcp login mirrors

login opens the browser to sign in; tokens refresh automatically.

On claude.ai or Claude Desktop:

Settings → Connectors → Add custom connector
Paste https://api.runmirrors.com/mcp
Click Add, then Connect

Connect opens the browser to sign in and approve. Works on free and paid plans.

Every other client (Zed, Warp, Goose, Cline, Windsurf, …), the one-click installs, and headless/CI setups with a workspace key are on the MCP server page.

Install the collector

The collector streams your agent’s real production traces to Mirrors, the raw material every environment is built from. First mint a workspace API key (mk_live_…) at runmirrors.com under Settings → API keys, then install the collector and initialize it before your agent runs. It auto-instruments LangChain/LangGraph, Anthropic, and OpenAI, ships traces in the background, and never raises into your app.

# pip install mirrorkit
import mirrorkit
mirrorkit.init(api_key="mk_live_...", project="my-agent")

// npm install mirrorkit
import * as mirrorkit from 'mirrorkit';
mirrorkit.init({ apiKey: 'mk_live_...', project: 'my-agent' });

// go get github.com/ai-singhal/mirrorkit-go
mirrorkit.Init(mirrorkit.Options{APIKey: "mk_live_...", Project: "my-agent"})

Run your agent normally and traces start streaming. Details per language: Python, TypeScript, Go.

Build your first environment

Turn the collected traces into an environment: a schema, a seeded database, and bound tools, a runnable copy of your agent’s world, scored per tool for fidelity. Ask your MCP client, or use the CLI:

Build a mirror from my "my-agent" project and call it my-agent.

pip install "mirrorkit[cli]"
mirrors login                                   # paste your mk_live_… key
mirrors build --project my-agent --name my-agent

You can also build from the dashboard under Ingest → Build. Prefer files over a live stream? mirrors build traces.jsonl --name my-agent ingests a trace log directly.

Create your first eval set

An eval set is the suite of sessions replayed against every change. Auto-author cases from your real traces, then keep the ones that matter:

Generate an eval set for my-agent and save it as "smoke".

mirrors eval generate my-agent --save-as smoke

Hand-write or extend cases with mirrors eval create my-agent --name smoke --from cases.json. See How it works for what a case scores.

View your first result

Run the eval set and read the result: per-tool coverage and accuracy, plus any session that would have broken.

Run the "smoke" eval set for my-agent and show me the result.

mirrors eval run <eval-set-id>     # kick off the run
mirrors run show <run-id>          # inspect scores + traces

Or open the run in the dashboard. From here, wire the same eval set into CI so every PR that touches your agent replays it. See CI gate.

Next steps

How it works

Fidelity scoring, deterministic seeding, business context, and evals.

MCP server

Every client, one-click installs, and headless setups.

CLI

The mirrors command: full parity with the web app.

CI gate

Block regressions at the pull request, Vercel-style.

Get started

Collect traces

Drive Mirrors

Account

Resources

Quickstart: test an AI agent

How do I set up Mirrors, step by step?

Next steps

How it works

MCP server

CLI

CI gate

​How do I set up Mirrors, step by step?

​Next steps

How it works

MCP server

CLI

CI gate

How do I set up Mirrors, step by step?

Next steps