AI eval and benchmark tools
Evaluation tools that score agent outputs, catch regressions and compare prompts and models. Of the 121 tools on our evaluating agents page, 42 fit this narrower job. By pricing: 25 open source, 3 source-available, 8 free or freemium, 3 paid, usage-based or enterprise, 3 pricing on the maker's site. The most-starred open repos here are Cua (29,966 stars), MLflow (28,418 stars), Opik (22,700 stars). Most come from the evals & observability (27), frameworks (5), browser/computer use (2) categories.
Updated
Cua
Open-source infrastructure for computer-use agents: sandboxes, drivers across operating systems and benchmarks.
Added
MLflow
Open-source platform for tracing, evaluating and managing agents, LLM apps and ML models.
Added
Opik
Comet's open-source platform for tracing, evaluating and monitoring LLM apps, RAG systems and agent workflows, self-hosted or in the cloud.
Added
Arize Phoenix
Open-source tracing and evaluation for LLM apps.
Added
Oumi
Open platform for fine-tuning, evaluating and deploying open agentic LLMs and VLMs.
Added
Agents CLI
Google's CLI and skills that teach coding assistants to build, evaluate and deploy agents on Google Cloud and the Gemini platform.
Added
SWE-bench
Benchmark that tests whether language models and coding agents can resolve real GitHub issues, with multimodal and verified variants.
Added
Apache Maka
Incubating Apache agent harness and workspace that records everything it does and publishes benchmarked run results.
Added
Kiln
Free desktop app and open-source library for building AI systems with evals, RAG, agents, fine-tuning and synthetic data in one workbench.
Added
Agenta
Open-source LLMOps platform for prompt management, evaluation and observability.
Added
fast-agent
Framework to build, run and evaluate agents with strong MCP, ACP and A2A support.
Added
TruLens
Evaluation and tracking for LLM experiments and AI agents.
Added
EvoAgentX
Framework for building, evaluating and automatically evolving LLM agents and agentic workflows.
Added
OSWorld
Benchmark for multimodal computer-use agents on open-ended tasks in real desktop environments.
Added
OpenLIT
Open-source platform to trace, evaluate and guard agents and LLM apps.
Added
Qwen-UI-Agent
Tongyi foundation GUI agent model that operates mobile, computer and browser interfaces, trained and evaluated on real devices.
Added
MCPJam Inspector
Open-source testing and evals platform for MCP servers, MCP apps and ChatGPT apps across many clients and models.
Added
Future AGI
Open-source platform for evaluating, tracing, simulating and guarding LLM and AI agent applications, self-hostable or cloud.
Added
tRPC-Agent-Go
Go framework for production agent systems with graph workflows, tools, memory, A2A, AG-UI, MCP, evaluation and OpenTelemetry.
Added
OpenAPPA
OpenAPPA is a deterministic guardrail that sits between an AI agent and its tools and tracks how data flows across tool calls, instead of asking a second model to judge each call. Policies decide whether untrusted or sensitive data may reach a given action, which blocks data exfiltration caused by prompt injection or hallucination while letting agents finish their tasks. It plugs into Claude Code, Pi, OpenCode, kAgent and Archestra, comes with reusable policy "batteries", validation, observability and reporting, and is backed by a NeurIPS workshop paper with published benchmarks. It is in preview as an open RFC.
Added
AIDE ML
Tree-search agent that drafts, debugs and benchmarks machine-learning code until a target metric is optimized.
Added
BrowserGym
ServiceNow's Gym environment for building, testing and evaluating web agents across browser benchmarks.
Added
AI Beat
An open-source AI red-teaming suite: PromptBeat attacks models and prompts, and AgentBeat runs black-box security evaluations against HTTP-exposed agents.
Added
Tracely
Tracely is open-source, trace-native CI/CD for AI agents: it grades every production trace with evaluators, clusters failures into issues, freezes bad runs into replayable test cases and blocks pull requests that would reintroduce them.
Added
Agent Memory Leaderboard
Public leaderboard and open evaluation protocol, run by a group of universities and labs, for comparing long-term memory systems for agents in recurring submission rounds.
Added
Rogue
Agent evaluator and red-team tool that probes your agent with test conversations and attacks and reports pass or fail against your policies.
Added
Tov
Tov is an experimental programming language designed to be written by coding agents. It uses TypeScript syntax minus a few traps (no any, null or loose equality, separate int and float types, and failing calls marked with try), so agents already know how to write it, and its compiler returns machine-readable fixes the agent can apply after every edit. Programs compile to small, fast native binaries, with published benchmarks against Rust, Bun and Node. Agents can learn it from the site's llms.txt.
Added
FiftyOne MCP Server
Voxel51's MCP server and skills that let AI agents load, explore, curate and evaluate computer vision datasets and run FiftyOne operators and the App from natural language.
Added
Arize AI
Agent observability and evaluation platform with tracing, evals and experiments for improving AI agents.
Added
Braintrust
Evals and observability platform for AI products.
Added
Confident AI
Enterprise platform for AI evals and observability that gives teams one standard for measuring and monitoring AI quality.
Added
CoreWeave Forge
One product to run, observe, evaluate and post-train agents (W&B, OpenPipe, sandboxes).
Added
Coval
Testing and evaluation platform for voice agents, with simulations, production evals and human QA.
Added
Giza
Self-custodial autonomous agents that evaluate markets and execute on-chain capital strategies (Giza World, ARMA).
Added
Hume AI
Voice AI company focused on emotionally intelligent speech, with voice-native evaluation and human judgment tooling for voice agents.
Added
Kore.ai Autoloop
Optimization engine in Kore.ai Agent Platform Artemis that evaluates agents against business goals, diagnoses failures from traces, applies targeted fixes, and re-verifies; GA Oct 2026.
Added
Langfuse
Open-source LLM observability, tracing and evals.
Added
LangSmith
Tracing, evals and monitoring for LLM apps and agents.
Added
Openlayer MCP
Openlayer's MCP server, letting coding agents pull evaluation results, failing tests and monitoring data from Openlayer so they can debug and improve AI systems from the IDE.
Added
Respan
Observability and evaluation platform for LLM apps and agents.
Added
W&B Weave
Weights & Biases toolkit for tracing and evaluating agents.
Added
YRIKKA Text-to-Model
YRIKKA's ORION workflow that turns a plain-language mission description and a few unlabeled images into generated training data, candidate models and an evaluated, selected vision model.
Added
GitHub repos (28)
| Repo | GitHub stars | Licence | Last push |
|---|---|---|---|
| Cua ↗ | 29,966 | MIT | 2026-10-21 |
| MLflow ↗ | 28,418 | Apache License 2.0 | 2026-10-21 |
| Opik ↗ | 22,700 | Apache License 2.0 | 2026-10-21 |
| Arize Phoenix ↗ | 11,828 | Elastic License 2.0 (source-available) | 2026-10-21 |
| Oumi ↗ | 9,400 | Apache License 2.0 | 2026-10-21 |
| Agents CLI ↗ | 6,137 | Apache License 2.0 | 2026-10-20 |
| SWE-bench ↗ | 6,067 | MIT | 2026-09-02 |
| Apache Maka ↗ | 5,842 | — | 2026-10-21 |
| Kiln ↗ | 5,195 | — | 2026-10-21 |
| Agenta ↗ | 4,840 | MIT License, except ee/ directories (separate licence) | 2026-10-20 |
| fast-agent ↗ | 3,935 | Apache License 2.0 | 2026-10-20 |
| TruLens ↗ | 3,602 | MIT | 2026-10-19 |
| EvoAgentX ↗ | 3,385 | — | 2026-08-27 |
| OSWorld ↗ | 3,217 | — | 2026-10-21 |
| OpenLIT ↗ | 2,853 | Apache License 2.0 | 2026-10-21 |
Top 15 of 28 by stars. Stars and licence come from each repo, checked when the listing was last verified.
FAQ
What are the best AI tools for evaluating agents focused on evals and benchmarks?
We list 42. The first three by GitHub stars and featured status are Cua: Open-source infrastructure for computer-use agents: sandboxes, drivers across operating systems and benchmarks.; MLflow: Open-source platform for tracing, evaluating and managing agents, LLM apps and ML models.; Opik: Comet's open-source platform for tracing, evaluating and monitoring LLM apps, RAG systems and agent workflows, self-hosted or in the cloud.
Are there free or open-source options?
Yes: 25 open source, 3 source-available, 8 free or freemium, 3 paid, usage-based or enterprise, 3 pricing on the maker's site. Pricing is as each maker states it.
More from evaluating agents
Tools are grouped by what each says it does. Missing one? Submit it free.