terminal-bench 开源项目推荐（101 个）

harbor-framework /harbor

评估与改进智能体的框架

itayinbarr /little-coder

针对较小的LLM进行优化的安全带

zli12321 /LHTB

长视线终端基准测试及密集奖励分级

simple-agent-lab /RSIHub

A research framework for principled agent self-improvement under frozen evaluators and declared mutation boundaries, recording verifiable lineage to make it reproducible and auditable.

tribgames /mixdog

Open-source desktop coding agent where a lead model plans and cheaper models handle search, edits and review.

alanhuangyoo /crux

Taking the pi coding agent to Claude Code-level performance: 0.539 → 0.773 pass@1 on Terminal-Bench 2.1 with the same self-hosted 27B model, by re-engineering the agent.

UiPath /coder_eval

Playwright for coding agents. Test that your skills, MCP servers, and CLIs actually work when an agent uses them — sandboxed YAML suites, A/B experiments, CI gates.

LiberCoders /CLI-Gym

《CLI-Gym：通过智能体环境反转实现可扩展 CLI 任务生成》官方实现

mmprotest /villani-code

Some tools help. Some tools assist. Villani Code intervenes.

todoforai /edge

Open-source AI agent that runs on your own computer and does your tasks. Any model (Claude, GPT, Gemini, Grok, Ollama), drives your real browser, shell, files, DOCX/XLSX.

strands-labs /benchmark-harnesses

Strands-based agents and harnesses for agentic benchmarks.

simple-agent-lab /simple-long-horizon-agent

Simple Long Horizon Agent - A simple yet effective AI agent for learning, experimentation, and long horizon work.

quqxui /RSI-Flow

A recursive self-improvement (RSI) workflow that evolves both target agents (e.g., Pi and DeepSeek Harness) and the RSI system itself.

macanderson /stella

A coding agent that builds it's own tools and mines it's own skills and always ships with receipts.

plaume8 /spoox

Spoox CLI - Terminal Agent - SPlit lOOp eXand agent

scitix /Agent-Sandbox

Fast, Multi-Cloud Sandbox Engine for AI Agents

aktasbatuhan /compound

Which host should serve your open model? Run your workload across pinned, verified providers and get one table: success, cost, latency, cache hits.

timwhitez /dsh-self-evolving

Evidence-first, crash-resumable self-evolution engine for DeepSeek Harness and Harbor.

bugyal /skeptic

Runs BOTH controls (nop + oracle) across EVERY task in a coding-agent benchmark, classifies the outcome, captures evidence, and lints leaks statically. One binary, zero LLM calls.

li-boxuan /Terminal-bench-OpenHands-trajectories

Trajectories for running OpenHands on Terminal Bench