Browsentic

Evidence shows a fairly detailed permission and confirmation design: approvals.md gives a rule table with most-severe-wins, MCP external callers default confirm to deny, credentials are sealed before return, and blocked sites are enforced extension-side. least_privilege scores only 2: the extension injects into every page, the Bridge can spawn any logged-in CLI, and SECURITY.md admits any local process running as the user can drive a paired browser, so the permission surface is broad. data_flow_transparency scores 2: README and SECURITY describe loopback-only, no telemetry, no cloud, but lack verifiable detail on the agent CLI's own data egress. sensitive_data_handling scores 2: sealing and 0600 file modes are described, but SECURITY.md calls the sanitizer a filter not a proof, and held secrets sit in extension session storage in the clear. dependency_security scores only 1: package.json shows yarn 4.17.1 and scripts, with no lockfile audit, dependency inventory, or vulnerability-scan evidence in the provided files. external_effects scores 2: downloads, uploads and form submissions are gated, but there is no compensation story for side effects already executed. rollback scores only 1: uninstall and approvals clear exist, but no rollback path for irreversible actions such as submitted forms or sent messages. source_attribution scores only 1: LICENSE carries a copyright line, but publisher identity is unverified and no NOTICE or third-party attribution list appears.

self_consistency scores 2: README, install, configuration and approvals broadly agree on ports and protocol versions, but README claims 52 tools while the install sample shows 0.8.0 and package.json shows 0.9.0, a version drift. dependency_availability scores only 1: docs say src/daemon is a separate Yarn project with Yarn 4 vendored, but the provided files contain no lockfile or dependency manifest, so availability cannot be statically confirmed. failure_messages scores 3: troubleshooting.md maps symptoms to error codes, causes and fixes, and an errors reference is cited; coverage and specificity are strong for static material.

audience_and_scenarios scores 2: it targets both ordinary users and developers across many browsers and CLIs, but scenario descriptions lean marketing. capability_boundaries scores 3: limits.md, the nine Firefox-missing tools, experimental Windows and Android, and the non-configurable ceiling table are all explicit. trigger_precision scores 2: rules and scope derivation are defined, but scope is deliberately unconfined on a blank tab, a loose default that leaves a precision gap. environment_fit scores 2: macOS/Windows/Linux and many browsers are supported, but Windows and Android are experimental and Firefox is feature-limited, so fit is uneven.

information_architecture scores 3: docs/guide, docs/internals and docs/reference are cleanly layered with a complete README index. install_notes scores 3: install.md covers platforms, browsers, unpacked, source builds, updates and uninstall with concrete steps. naming_stability scores 2: commands and config keys are consistent, but there is an MIT-to-Apache-2.0 license change and version numbers disagree across files. examples_and_faq scores 2: config examples and a troubleshooting table exist, but there is no standalone FAQ and few end-to-end examples. known_limitations scores 3: limits.md and SECURITY.md plainly list uncovered areas and risks. license scores 3: LICENSE is the full Apache-2.0 text and package.json agrees. versioning_changelog scores only 1: package.json has a version, but no CHANGELOG or release notes are in the provided material. maintenance_responsibility scores only 1: SECURITY.md describes a one-maintainer project, publisher identity is unverified, and ownership of maintenance is unclear.

output_usability scores 2: the side panel shows steps, approval cards and results, so usability is designed in, but static material cannot verify actual output quality. marginal_value scores 2: reusing an existing CLI subscription with no API key is a clear increment, yet it overlaps existing browser-automation tools, so the increment is bounded. cost_benefit scores 2: free, open source and local-first lower cost, but installing a Bridge, extension and CLI plus security review is non-trivial effort.

claim_traceability scores 2: most claims trace to specific doc sections and rule ids, but README badges and store links cannot be checked statically. cross_source_corroboration scores only 1: README, SECURITY and approvals corroborate some security claims, but there is no code, test or third-party audit backing. fact_inference_separation scores 2: docs separate implemented mechanisms from non-guarantees (e.g. prompt injection only mitigated), but marketing phrasing is mixed with factual statements.