x-use

Trust: least_privilege scores 2 — config docs cap browser sessions (1-16), tool timeouts, daily action caps, and media downloads are restricted to public HTTPS hosts with redirects disabled and no account cookies; deducted because dependencies include selenium, undetected-chromedriver and selenium-stealth (anti-detection/automation surface) widening the privilege footprint. user_confirmation scores 2 — draft_mode defaults on, messages and follows always require individual approval, approve_draft is explicit; deducted because direct publishing mode, opt-in auto_drain and the legacy Selenium batch runner are separate execution paths that can bypass draft review. data_flow_transparency scores 2 — docs describe X_USE_HOME storage, LLM key precedence, env overrides and that keys are never logged; deducted because AI-client and optional model-provider data handling is only described as depending on the chosen client/provider. sensitive_data_handling scores 2 — cookie import, inbox PIN via env var consumed once, proxy credentials masked in tool responses, atomic backups; deducted because cookies are imported from a plaintext local JSON export with no encryption-at-rest discussion. dependency_security scores 2 — CI runs pip-audit against the locked dependency set, uv.lock check and pip check; deducted because several runtime deps (selenium, webdriver-manager, fake-headers, openai, requests) are unpinned and scanning only covers published advisories. external_effects scores 1 — tools post, reply, DM, follow, delete, unfollow and manage conversations, i.e. real external side effects; despite drafts and budgets, direct publishing, queue, legacy batch and community engagement paths can produce irreversible external actions, and uncertain outcomes must be manually inspected by action ID. rollback scores 1 — persistent drafts, queue retries and resolve_action_outcome exist, but published posts, sent messages and follows have no automatic rollback, only manual deletion/reversal. source_attribution scores 1 — README states MIT, author Huzaifa Shoukat, independent community project, former name twitter-automation-ai, but publisher identity is unverified by the FollowAgents registry (unknown, not suspicious); attribution rests on self-description.

Reliability: self_consistency scores 2 — pyproject reads version dynamically from src/xuse/__init__.py so package, wheel and serverInfo cannot disagree, and README/config docs broadly agree on draft_mode, daily caps and tool counts; deducted because README claims 82 tools in 3.2.0 while CONFIG_REFERENCE is still framed around 3.1.0, a version-narrative mismatch. dependency_availability scores 1 — dependencies are declared but several are unpinned, and setup additionally requires Patchright Chromium plus system libraries, a heavy environment dependency; CI covers multiple platforms but static files cannot prove installability. failure_messages scores 2 — code and tests show structured errors (BrowserActionError, BrowserBlocked with reasons like login_required, challenge, account_locked, profile_mismatch) and config docs describe structured cold-start/tool timeouts; deducted because no complete user-facing error message catalogue is shown.

Adaptability: audience_and_scenarios scores 2 — README targets research, inbox management, reviewed publishing and outreach, with client examples for Claude/Codex/Cursor/Windsurf and a first workflow; deducted because it is single-platform (X) with limited scenario breadth. capability_boundaries scores 2 — docs explicitly list limits: browser reads bounded by visible content, opening conversations changes read state, encrypted inboxes need a PIN, analytics only reports the Premium gate, community posting unverified in the async runtime; boundaries are well stated. trigger_precision scores 1 — 82 tools across multiple execution paths (draft, direct publish, queue, auto_drain, legacy batch) mean triggering depends on configured policy; docs point to the config reference but provide no clear trigger decision matrix. environment_fit scores 2 — requires Python 3.11+ and Chromium, CI covers Windows/macOS/Linux and 3.11-3.14, stdio config and uvx usage are given; deducted because desktop clients need absolute paths and X_USE_HOME env vars, raising setup friction.

Convention: information_architecture scores 2 — README has clear sections, capability table, FAQ and doc index links, with docs split by topic (config, controls, runtime, workflows); deducted because several key docs are only linked and not present in the evidence. install_notes scores 2 — pip install, Patchright Chromium install, X_USE_HOME setup, MCP client JSON config, source install and uvx usage are all provided. naming_stability scores 1 — the project was renamed from twitter-automation-ai to x-use, package name x-use-mcp coexists with x-use/x-use-mcp executables, and 3.0.0 vs 3.2.0 client commands differ, creating migration burden. examples_and_faq scores 2 — README has a first workflow example, a multi-question FAQ, workflow doc links and presets; deducted because examples are mostly descriptive rather than complete copy-paste scripts. known_limitations scores 2 — README and config docs explicitly list browser read bounds, read-state changes, PIN, analytics limits, platform limits and unverified legacy engine. license scores 2 — LICENSE is full MIT text, pyproject declares license=MIT, README and badge agree. versioning_changelog scores 1 — README references a CHANGELOG and version progression (3.0.0→3.1.0→3.2.0), but CHANGELOG content is not in the evidence and config docs lag in version, deducted. maintenance_responsibility scores 1 — README names maintainer Huzaifa Shoukat with contribution channels and SECURITY.md gives a reporting email, but publisher identity is unverified by the registry and SECURITY.md explicitly disclaims LTS or response-time guarantees.

Effectiveness: output_usability scores 2 — tools return structured results (e.g. analytics snapshot with status, metrics, coverage, observed_at), drafts expose exact text and ID, media attaches as MCP content blocks; deducted because some outputs depend on client capabilities (vision, image support). marginal_value scores 2 — no X API key, multi-account, reviewed publishing and local outreach management offer clear incremental value for specific users; deducted because similar browser-automation MCP projects exist and reliance on anti-detection libraries carries platform risk. cost_benefit scores 1 — requires Chromium install, cookie import and data-directory setup, and understanding multiple execution paths and safety budgets, a high learning/ops cost; platform layout changes, expired sessions and account challenges can interrupt work, making payoff uncertain.

Verifiability: claim_traceability scores 1 — most README capability claims map to config docs and test files (e.g. analytics paywall tests, chat reuse tests), but quantitative claims like 82 tools, seven prompts and four resource patterns are not itemized in the evidence, and key docs (CHANGELOG, MCP_GUIDE, TOOL_VALIDATION) are not present. cross_source_corroboration scores 1 — README, CONFIG_REFERENCE, pyproject, CI and tests partially corroborate each other on draft mode, dependencies and CI matrix; deducted because version numbers (3.1.0 vs 3.2.0) and tool counts are inconsistent and no independent third-party corroboration exists. fact_inference_separation scores 2 — docs reasonably separate observed evidence from inference (analytics tests explicitly accept only observed evidence and never fabricate metrics; TOOL_VALIDATION separates local protocol checks, synthetic browser fixtures and authorized live-account checks), so little deduction.