Vessel Browser

The README describes visible supervision, scoped approvals, reject-and-steer, domain policies, undo, and session restoration; tests also cover approval states and runtime behavior, giving ordinary confirmation and rollback some concrete support. However, several safeguards are chiefly README claims, and auto mode can bypass actions explicitly marked as requiring approval, so confirmation and rollback do not receive full marks. The README describes encrypted conversation history, in-memory-only handling when secure storage is unavailable, redacted run records, and credential blind fill with domain limits and per-use consent. These details support sensitive-data handling and data-flow transparency, but the supplied material lacks code corroboration and a complete data-flow account. Dependency-audit scripts and CI audit steps are present, but no audit results are supplied, so dependency security receives partial credit. MCP, local AI providers, and browser actions can affect external services; the README describes navigation and form work alongside approval and domain controls, while evidence for least privilege is limited. package.json identifies an author, email, and repository, but publisher identity is unknown under the prompt, so attribution is only partial.

package.json and the README broadly align on Electron, Node, and the main features. The README makes broad platform claims while also calling Linux the most mature target and macOS source packaging available; installation guidance specifically recommends Linux AppImage and describes Windows installers, leaving uneven maturity and delivery paths. Node/npm versions, a lockfile-based CI install, and CI checks are specified, but Electron downloads depend on external availability; the README offers a mirror as a workaround, so dependency availability is only partly handled. Runtime health checks, error records, interrupted-run handling, and tests asserting concrete error responses support failure messaging. Full marks are withheld because user-facing failure handling is not comprehensively shown.

The README names ordinary browsing, research, form work, long sessions, and harnesses such as Hermes, OpenClaw, and MCP clients, giving useful audience and scenario coverage. It lists capabilities in detail and marks Research as beta, some features as Premium, and Linux as the most mature target; a complete external-tool permission boundary is still missing. Startup instructions, shortcuts, MCP connection options, and domain policies make use conditions fairly clear, but the approval requirements for each action are not specified. Node 22, Linux installation, macOS builds, Windows installers, and an Electron mirror workaround are documented; differing platform maturity and macOS signing/notarization limitations prevent a full environment-fit score.

The README is organized into quick start, features, architecture, installation, and external-agent setup. It covers AppImage, npm, source, development, and packaging, but the source install pipes a remote script to a shell without a verification method in the supplied material. Package, command, MCP, and version names are mostly consistent. The README provides commands and Hermes/OpenClaw setup steps, but no complete FAQ; the supplied README text also ends mid-section. It discloses early development, Research beta status, platform maturity, lack of notarization, and that interrupted runs are not automatically resumed. The MIT license text is complete and earns full marks. package.json gives version 0.1.0, but no changelog or clear release history is included. An author and contact email are listed, while maintenance responsibility, support commitments, and update ownership remain unclear.

The README describes usable outputs and controls for long sessions, approvals, run records, bookmarks, rollback, and report export; tests also show structured action events and error responses. These support practical output usability, though no complete end-user output example is provided. Visible browser supervision, durable state, and MCP control offer a clear additional capability for agent browsing, but many advanced features require Premium and the product is early software. The README identifies free and paid features and documents installation effort, but does not quantify resource use, service costs, or maintenance burden, limiting cost-benefit evidence.

README claims about installation, architecture, protocol endpoints, retention, and approvals can be traced to package.json, CI configuration, and some tests. Many feature claims remain README-only in the supplied material. The README, configuration, and tests provide limited cross-source corroboration; CI describes intended checks but does not establish that this revision passed them. The README distinguishes implemented features from beta status, Premium gates, and platform maturity, but promotional claims are not consistently separated from independently supported facts.