Test your AI agents. Catch failures before your users do.

---

Quick Links: Quickstart · Assertions · Failure Types · Contributing

---

pip install playagent[all]

from playagent import record
from playagent.adapters.openai import OpenAI

client = OpenAI()

@record
def run_agent(user_input: str):
    return client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": user_input}],
    )

  session   sess_a1b2c3d4
  agent     run_agent
  started   2026-04-05 09:14:32
  duration  3.24s
  status    passed

──────────────────────── turn 1 ─────────────────────────
  model     gpt-4o
  latency   812ms

  ▸ user
    What's the weather in Lagos today?

  ▸ assistant
    I'll look that up for you.
    ⬡ tool call  get_weather
      location  "Lagos, NG"
      units     "celsius"

Command	What it does
playagent trace list	Lists recent trace sessions.
playagent trace view <trace_id>	Shows turn-by-turn trace details.
playagent report	Shows aggregate pass/fail and failure breakdowns.
playagent report --format json	Emits report stats as JSON for CI pipelines.
playagent --version	Prints installed PlayAgent version.

Why PlayAgent

• You stay local-first. PlayAgent writes to SQLite on your machine; nothing is sent to a hosted dashboard by default.
• You can test behavior, not only outputs. Assertions check tool-call order, parameters, and call counts directly.
• If you already use LangSmith, PlayAgent is a smaller option for local SDK-level checks; if you need hosted traces, collaboration, and observability dashboards, LangSmith is the better fit.

License

MIT