WeaveBench
Copyright (c) 2026 Wanli Li and contributors

This product includes adaptations and integrations of third-party software
distributed under their respective licenses. WeaveBench itself is released
under the MIT License (see LICENSE). The following third-party components
are referenced or required at runtime; their copyright remains with their
respective owners.

1. OSWorld (https://github.com/xlang-ai/OSWorld) — MIT License.
   The contents of `weavebench/desktop_env/` are derived from the OSWorld
   `desktop_env/` Python package. We bundle the `controllers/`, the `docker/`
   KVM provider, the `providers/base.py` abstraction, the `actions.py`
   keyboard table, the `server/` in-VM Flask shim, and the top-level
   `desktop_env.py` entry point — with two modifications:
     (a) the OSWorld `evaluators/` subpackage is *stripped*, since WeaveBench
         tasks embed their own `## Grader` Python block and do not need
         OSWorld's evaluator pipeline. A small inlined `compare_urls` helper
         keeps `controllers/setup.SetupController` functional.
     (b) optional third-party imports (`playwright`, `pydrive`, `dotenv`,
         `requests_toolbelt`, `aws.proxy_pool`, …) are wrapped in
         `try/except ImportError` so missing optional deps no longer break
         `import`.
   Cloud providers (`aws`, `azure`, `aliyun`, `gcp`, `virtualbox`, `vmware`,
   `volcengine`) are also stripped — only `docker` (KVM/qcow2) and our
   `docker_lite` (headless container) are bundled. Install upstream OSWorld
   separately if you need any of the other providers.
   We additionally contribute the `providers/docker_lite/` subpackage which
   adds a fully headless (no X server) Docker variant used for CLI ablation.

2. WildClawBench (https://github.com/internlm/WildClawBench) — the
   four harness adapters (`openclaw_agent.py`, `codex_agent.py`,
   `claudecode_agent.py`, `hermes_agent.py`), the GUI plugin under
   `assets/computer_tool_plugin/`, the wire-layer patches under
   `assets/{openclaw,codex,claudecode,hermes}_patches/`, and the
   trajectory-aware judge under `eval/agent_judge/` are derived from
   WildClawBench's harness staging ground.

3. OpenAI Codex CLI (https://github.com/openai/codex) — Apache-2.0.
   `codex.tar.gz` repackages the `@openai/codex` npm package + Node runtime
   for in-VM use; the original license terms apply.

4. Claude Code (https://github.com/anthropics/claude-code) — see upstream
   terms. `claudecode.tar.gz` repackages the `@anthropic-ai/claude-code` npm
   package for in-VM use.

5. Hermes (Nous Research, https://hermes-agent.nousresearch.com) — see
   upstream terms. `hermes.tar.gz` repackages the published Hermes 0.14.0
   wheel + its Python 3.11 interpreter for in-VM use.

6. Model Context Protocol (modelcontextprotocol.io) — MIT License.
   `hermes_mcp_wheels.tar.gz` is a snapshot of `mcp==1.27.1` and its
   transitive dependencies, redistributed unchanged for offline install.
