Alternatives to Kashikoi

CodeClash benchmarks goal-oriented software engineering by making AI models build and evolve codebases to win 15-round arena tournaments.

Goodfire's Silico is an AI interpretability agent for reverse-engineering, debugging, and steering model internals.

Sharbo builds a skill network for AI agents — encode expert workflows once and let every agent activate them autonomously.

AppWorld benchmarks interactive coding agents on 750 multi-app tasks across 457 APIs.

Build custom RL environments in code and run agent evaluations with trace QA on every run

Open source AI engineering platform for agent and LLM tracing, evaluation, prompt management, and ML lifecycle work.

Fiddler AI is an enterprise control plane for agent observability, inline guardrails, and auditable AI governance.

Synth is a research workbench and hosted compute for optimizing coding agents with reproducible, evidence-backed evals.

PromptLayer is the collaboration layer for AI engineering teams — a prompt CMS, eval harness, and agent observability stack in one dashboard.

Free Hugging Face agent leaderboard ranking LLMs on planning, tool use, and multi-step agentic tasks

Patronus AI builds RL environments and training data so frontier agents learn judgment on the job

Arena AI ranks chat models and coding agents on live head-to-head human votes and real agent sessions.

Open-source benchmark that scores AI agents on real, multi-step software-company work tasks.

Compare 400+ AI models on one LLM Stats Score blending benchmarks, output speed and API pricing.

VisualWebArena is a 910-task benchmark from Carnegie Mellon that scores multimodal browser agents on whether they actually complete visual web tasks.

PandaProbe is open-source agent observability that traces runs, evaluates outcomes, and turns failures into validated fixes via its Repair

OpenCompass (司南) benchmarks LLMs, VLMs, and AI4S models on dated, dimension-by-dimension leaderboards

Free AI benchmark where models build 3D voxel structures from text prompts and humans vote on spatial reasoning.

FFMPerative Outrider turns AI research into review-ready draft pull requests matched to your codebase, then measures whether they helped.

BALROG is an open ICLR-published benchmark that scores agentic LLM and VLM reasoning across seven procedurally generated games.

Open-source LLM observability: trace agents, manage prompts, and run evals on one MIT-licensed platform.

Comet's Opik brings LLM observability and evaluation that turns agent traces into code fixes

Automated QA for voice and chat AI agents — simulate conversations, monitor live calls, and close the loop from failing test to patched prompt.

Put a hard question to three frontier models — they answer blind, cross-examine each other, and hand back the one answer that survived, with a verdict.

Trace, replay and debug AI agent runs — LLM calls, tool use and multi-agent interactions in one timeline, via a single SDK

Conan is a native macOS HUD for Claude Code that streams every prompt, tool call, skill and token live as the session runs.

ViDoRe Benchmark: an open visual document retrieval benchmark and ColPali model suite for enterprise RAG evaluation.

Test hundreds of AI agent versions at once with RapidFire AI's parallel optimization engine

Weights & Biases tracks ML experiments, versions models and datasets, and traces LLM apps — now inside CoreWeave Forge

AgentNotch streams live Claude Code and Codex token and cost telemetry into your Mac's notch