litmuseval 0.2.1

The open-source evaluation framework for AI agents — test, compare, ship with confidence

LitmusAI

Test framework for AI agents. Think pytest for LLMs — assertions, cost tracking, safety scanning.

pip install litmuseval

Why I built this

I was switching between GPT-4.1 and Claude, tweaking prompts, changing architectures — and I had no idea if anything was actually getting better. My "evaluation" was typing questions into a chat window and eyeballing the responses.

So I built this. Now I can run the same tests across models and know exactly what's working.

Quick start

import asyncio
import litmusai
from litmusai import Agent, TestSuite, TestCase, Numeric, Contains, evaluate

litmusai.configure(api_key="sk-...")

agent = Agent.from_openai_chat(model="gpt-4.1")

suite = TestSuite(name="basics")
suite.add_case(TestCase(
    id="math", name="Percentage",
    task="What is 15% of 240?",
    assertions=[Numeric(36, tolerance=0.01)],
))
suite.add_case(TestCase(
    id="fact", name="Author",
    task="Who wrote 1984?",
    assertions=[Contains(["Orwell"])],
))

results = asyncio.run(evaluate(agent, suite))
# ✅ 2/2 passed | $0.0003 | 937ms avg

Or write tests in YAML:

name: my-tests
cases:
  - id: refund
    task: "I want to return my shoes"
    assertions:
      - type: contains
        patterns: ["return", "refund"]
        mode: any
      - type: not_contains
        patterns: ["sorry, I can't help"]

litmus run --suite my-tests.yaml --agent my_agent:agent

What it actually does

Assertions — 15 types, not just substring matching. Numeric extraction (handles "thirty-six"), JSON schema validation, embedding similarity, LLM-as-judge, regex, composable logic (All, AnyOf, AtLeast).

Safety scanning — 46 attack prompts across prompt injection, jailbreaks, PII leaks, bias, hallucination. My GPT-4.1 agent scored 78/100 — it literally replied "HACKED" to one of the injection tests.

Real cost tracking — costs come from actual API responses, not tiktoken estimates. Tiktoken can be off by 10-20%.

Multi-run stats — run the same test 5 times. Turns out some models pass a test 3 out of 5 times. You don't catch that with a single run.

Regression detection — diff two runs and see what changed. Pass rate up, cost down? Ship it.

Some numbers I found interesting

I ran the same suite across models:

Model	Pass Rate	Cost	Cost/Correct
GPT-4.1	100%	$0.017	$0.0034
Claude Sonnet 4	100%	$0.011	$0.0018
Claude Opus 4	83%	$0.043	$0.0085

Opus costs 14x more per correct answer than GPT-4.1 and scores lower. We were using it for months.

Connect your agent

Agent.from_openai_chat(model="gpt-4.1")           # OpenAI / compatible
Agent.from_azure(resource="r", deployment="d")     # Azure
Agent.from_function(my_fn)                         # any async function
Agent.from_url("http://localhost:8000/chat")       # HTTP endpoint
Agent.from_langchain(chain)                        # LangChain
Agent.from_crewai(crew)                            # CrewAI

Pipeline

Run eval + safety + report in one call:

import asyncio
from litmusai import Agent, Pipeline

agent = Agent.from_openai_chat(model="gpt-4.1", api_key="sk-...")

async def main():
    result = await Pipeline(
        agent, "coding",
        safety=True,
        runs=3,
        report="html",
    ).run()

asyncio.run(main())

Profiles

Presets for common scenarios:

litmus run -s coding -a agent:fn --profile quick       # fast iteration
litmus run -s coding -a agent:fn --profile thorough    # 3 runs, strict threshold
litmus run -s coding -a agent:fn --profile benchmark   # 5 runs, temp=0
litmus run -s coding -a agent:fn --profile ci          # strict threshold
litmus profiles                                         # see all

Custom profiles in YAML:

# .litmus/profiles/production.yaml
name: production
runs: 5
safety: true
safety_depth: thorough
threshold: 0.9
report: html

Built-in suites

8 suites, 50 test cases to start with. Not meant to be comprehensive — they're a starting point. Write your own for your domain.

litmus suites                                    # list them
litmus run --suite coding --agent my_agent:agent  # run one

coding · research · safety · planning · customer_support · summarization · instruction_following · tool_use

Custom assertions

from litmusai.assertions import Assertion, AssertionResult, register_assertion

class MaxWords(Assertion):
    def __init__(self, limit: int):
        self.limit = limit

    def check(self, response: str, **kwargs) -> AssertionResult:
        count = len(response.split())
        return AssertionResult(
            passed=count <= self.limit,
            score=min(1.0, self.limit / max(count, 1)),
            reason=f"{count} words (max {self.limit})",
            assertion_type="MaxWords",
        )

register_assertion("max_words", MaxWords)
# Now usable in YAML: { type: max_words, limit: 100 }

CLI

litmus run --suite coding --agent my_agent:agent    # evaluate
litmus run --suite tests.yaml --runs 5              # multi-run
litmus scan --agent my_agent:agent --level thorough  # safety scan
litmus diff --before run1.json --after run2.json     # compare runs
litmus report -r results.json --html report.html     # generate report
litmus init                                          # scaffold project

CI/CD

# .github/workflows/eval.yml
- uses: kutanti/litmusai@v1
  with:
    agent: my_agent:agent
    suite: coding
    threshold: 0.8

Development

git clone https://github.com/kutanti/litmusai.git
cd litmusai && pip install -e ".[dev]"
pytest                    # 729 tests
ruff check src/ tests/    # lint
mypy src/litmusai/        # types

~9K lines of code, 35 source files. MIT licensed.

License

MIT — Kunal Tanti