Open Computer Use (E2B cloud desktop)

The README describes the E2B Desktop Sandbox, keyboard/mouse/shell operation, multiple model providers, and a user pause-and-prompt feature. main.py opens a VNC client and attempts to terminate the sandbox in a finally block. This provides limited evidence of isolation, visibility, and a stop mechanism, supporting scores of 1 on those criteria. The supplied material does not show permission configuration, sensitive-data policies, per-action confirmation, or recovery procedures; sandboxing alone does not remove the external effects of agent actions. Dependency security is also only supported by version ranges in pyproject.toml, with no lockfile or review evidence, so it scores 1. The README credits and links E2B, model providers, and related projects, earning 2 for source attribution; maintainer identity and update responsibility remain unclear.

main.py prints agent errors and returns to the prompt, but offers limited handling or reporting for failures during setup, so failure_messages scores 1. The README's live-display description does not fully match the VNC flow currently started in main.py, and a code comment identifies an ffmpeg limitation for a particular desktop image; these visible inconsistencies keep self_consistency at 1. The README lists Python, Poetry, ffmpeg, E2B, and model-key requirements, but provides no evidence about availability or alternatives, so dependency_availability scores 1.

The README describes natural-language desktop operation, model swapping, several providers, and a startup example. That gives ordinary audiences and scenarios some coverage, earning 2 for audience_and_scenarios. It does not adequately define permitted actions, stopping conditions, or high-risk task boundaries, so capability_boundaries and trigger_precision each score 1. Although the README claims the design can work with different operating systems, the installation uses macOS brew and the code refers to Ubuntu and desktop-image behavior; evidence for environment fit is limited, scoring 1.

The README organizes features, design, model support, and setup, with links to configuration and provider files, earning 2 for information_architecture. Setup covers key prerequisites and commands but only demonstrates brew, without platform guidance or troubleshooting, so install_notes scores 1. Project, package, and script names are mostly consistent in the supplied files, earning 2 for naming_stability. There is a short startup example but no FAQ, so examples_and_faq scores 1. The README and code comments reveal some limitations, but there is no comprehensive known-limitations section, so known_limitations scores 1. LICENSE contains the full Apache 2.0 terms, earning 3. pyproject.toml states version 0.1.0, but no changelog is included in the supplied material, so versioning_changelog scores 1. The README invites provider-change PRs, but maintenance responsibility and update paths are otherwise limited, so maintenance_responsibility scores 1.

The product offers desktop operation, VNC viewing, model selection, and prompt input. The README provides startup commands, while main.py creates an output directory and handles Ctrl-C, giving output_usability a score of 2. Connecting models to a graphical desktop in a sandbox with visible screen access provides a clear capability beyond ordinary model chat, earning 2 for marginal_value. Cost-benefit evidence is limited to the required E2B and at least one model-provider API key; the supplied material gives no cost, resource-use, or performance information, so cost_benefit scores 1.

The README's setup requirements, provider list, and main operation descriptions can be partially checked against dependencies in pyproject.toml and the startup flow in main.py, earning 2 for claim_traceability. Cross-source support is limited, and the README's display-stream description differs from the VNC flow in main.py, so cross_source_corroboration scores 1. The documentation names concrete features and dependencies, but some security, cross-platform, and model-support claims are not sufficiently verified by the supplied code; the materials also do not clearly distinguish implemented behavior from design intent, so fact_inference_separation scores 1.