- Reliability 9/20
- The README claims all 283 catalog tools have implementations, documents a local live sweep written to /tmp/premiere-mcp-bridge/live-tool-sweep.json, and ships scripts/live-tool-sweep.mjs, so a real-host per-tool verification path appears to exist; the architecture (Node server emits ExtendScript into a temp dir, CEP polls and evals it via CSInterface, results written back as JSON) is coherent and self-consistent. It also discloses known host gaps honestly: no scripting API to remove effects, caption text often unreadable via DOM, import_ae_comps deliberately unadvertised because a generic .aep import can wedge the bridge, render-queue status only meaningful with Adobe Media Encoder, UXP is experimental. However, this review is static: no cached server manifest, no source files, no CI workflows were supplied, so MCP init handshake, tools/list output matching the README table, and 283-tool counts cannot be confirmed; no committed unit tests covering key paths are visible; and reliability depends on external conditions (CEP debug setting, fixed /tmp path, specific Premiere builds). Under the static cap of 12, deducted to 9 for missing verifiable execution evidence.
- Security and permissions 8/20
- No red line: no Adobe/MCP credentials are requested, no real tokens appear in install examples, and the only network surface is opt-in-to-disable anonymous telemetry plus update checks with four documented opt-outs (panel checkbox, config.json flag, PREMIERE_MCP_TELEMETRY=0, DO_NOT_TRACK=1). The disclosed payload is limited to install id, OS, version, tool name, success/failure, duration, error code, and failing Zod field names, explicitly excluding project names, media paths, arguments, and results. Real risks remain: the catalog grants extremely broad host control, including delete_project_item, ripple_delete, clearContents, execute_extendscript, evaluate_expression, and inspect_dom_object, yet only advisory 'inspect first, mutate small' guidance is documented, with no mandatory confirmation or scoping for destructive tools; the bridge uses a fixed world-readable /tmp/premiere-mcp-bridge directory, a same-machine trust boundary; the installer enables Adobe CEP debug mode and the distributed MCPB/CEP archives are unsigned, requiring user trust in the download source. Main risks are visible but confirmation and scoping are incomplete, so 8.
- Maintenance 11/20
- Not archived, MIT licensed, versioned on npm with GitHub Releases, has update checking with in-panel upgrade prompts, and ships KNOWN_ISSUES, CONTRIBUTING, PRIVACY, TERMS, and SECURITY governance files; macOS is stated as validated while Windows is only smoke-tested via GitHub Actions, and 20 translated READMEs plus an Agent Skill indicate sustained investment. 578 stars corroborate activity but add no points. Missing evidence: no commit history, cadence, issue-response SLA, dependency-update automation, or concrete security-response channel is visible in the supplied material; the status of 19 open issues is unknown; the '283 tools as of Aug 31, 2026' claim cannot be verified; and the reproducible signing workflow depends on private certificate secrets, so outsiders cannot reproduce it. Active but with governance and version-consistency gaps: 11.
- Documentation 14/20
- Documentation is unusually layered: global npm install, --install-cep/--doctor, Claude Desktop .mcpb bundle, Codex and Claude Code plugin marketplaces, source install with macOS/Windows branches, and a generic mcpServers JSON; tools are grouped into themed tables with plain-language descriptions; there is a recommended first call (verify_premiere_connection), a premiere://config/get_instructions resource, a 'Real Limits' section, troubleshooting trees for both bridge and client failures, a telemetry field disclosure with opt-outs, update behavior, and explicit labels distinguishing production CEP from experimental UXP and unsigned MCPB/CEP artifacts. Deductions: no per-tool parameter schemas or return-shape examples (users are told to introspect), critical setup detail is offloaded to QUICKSTART.md which was not provided, self-consistency of numeric claims (283 all implemented, import_ae_comps unadvertised) rests on unverified source, and there is no cost or performance guidance. 14.
- Setup experience 12/20
- The mainstream-client path looks clear: two npm commands plus --doctor, restart Premiere, open Window > Extensions > MCP Bridge (CEP) and Start Bridge, then use Claude Desktop, Codex, Claude Code, or VS Code Copilot; ready-made mcpServers JSON, per-client plugins, and a one-click MCPB bundle are provided, with macOS and Windows branch scripts. But evidence is thin: the installer enables an Adobe CEP debug setting (a host-level side effect), the bridge directory must be exactly /tmp/premiere-mcp-bridge and match what the panel shows, and without executing anything the success rate of --install-cep on real machines is unverifiable; Windows is smoke-tested only; the UXP panel additionally needs developer mode. Troubleshooting steps such as verifying the absolute dist/index.js path and setting PREMIERE_TEMP_DIR indicate a fairly manual, fragile fallback. Under the static cap of 15, scored 12.