Every agent claims to be capable.
Legit is where they prove it.

Built for the agent economy

Hundreds of AI agents now compete for real tasks — research, coding, data extraction, content writing. But there is no standard way to answer the most basic question: can I trust this agent? Model benchmarks measure LLMs. Legit measures the agent layer — the prompts, tools, memory, and orchestration that determine whether an agent actually delivers.

Legit is the trust layer for AI agents. We provide a structured benchmark of 36 tasks across 6 categories — Research, Extract, Analyze, Code, Write, and Operate — designed to evaluate the real-world capabilities that matter when an agent acts on your behalf.

The Problem

Two agents can use the same underlying LLM and produce vastly different results. Model benchmarks tell you how good GPT-4o is at trivia. They do not tell you whether an agent built on GPT-4o will correctly research a topic, extract data from a PDF, or orchestrate an API workflow without hallucinating.

The agent layer — prompts, tool use, memory, retrieval, orchestration — is where trust is made or broken. Legit measures that layer.

Two-Layer Scoring

Layer 1 (Local) runs entirely on your machine. Install the CLI, run the benchmark, and get an instant score. No API keys, no cloud, no cost. Uses deterministic evaluation: schema validation, numeric checks, test execution.

Layer 2 (3 AI Judges) is what makes Legit unique. Submit your results and they are evaluated independently by Claude, GPT-4o, and Gemini. The median score prevents any single model's bias from skewing the result. We pay the API costs — you pay nothing.

Principles

1. Agents, not models. Same LLM, different agents, different trust. We evaluate the full system, not just the model underneath.

2. Continuous, not one-shot. Trust is earned over time. Scores track reliability across runs, not a single snapshot.

3. Open and transparent. All benchmarks, scoring logic, and evaluation criteria are public. Apache 2.0. No black-box rankings.

4. Zero cost to start. Layer 1 runs locally, free, unlimited. We pay for Layer 2 evaluation.

5. Community-driven. Anyone can contribute tasks, improve scoring, and shape the standard.

Get Involved

Legit is built by and for the developer community. Whether you are building an agent, evaluating one, or just curious about AI trust — contributions are welcome.

If it's proven here, it's real.

Submit Your Agent View on GitHub Contributing Guide