FFmpeg Local Video Editor

The tools use local FFmpeg and typed arguments, reject caller-supplied filter graphs, and do not run through shell strings. Inputs are not overwritten; replacing an existing output requires --overwrite, and dry runs, fingerprinted plans, and cleanup of partial output on interruption are described. Optional speech recognition runs locally, but model weights may be downloaded from Hugging Face or elsewhere, so sensitive-data handling and least privilege are not fully established. A private vulnerability-reporting path and CodeQL are documented, but the supplied material contains no dependency audit or vulnerability status. Publisher identity is unknown; package metadata naming an author does not verify identity.

The README, package metadata, CI configuration, and test fixtures are broadly consistent. CI is described as covering several operating systems, Python versions, and FFmpeg versions, but no actual results for this revision are supplied, so compatibility and self-check claims remain primarily repository claims. doctor reports missing or unknown capabilities, and tools are documented to return a reason on stderr with a non-zero exit; the failure-message implementation is not fully shown, so this does not merit full marks.

The repository names Claude Code, Cursor, Codex, and MCP clients and gives concrete workflows from probing and editing through platform delivery. It also clearly states boundaries: it does not detect faces or subjects, judge highlights by content, or choose a crop anchor for the caller. Trigger conditions and environment needs are documented, but there is no complete platform installation matrix; FFmpeg build features, optional fonts, and ASR models affect availability.

The README has clear sections, a contents list, tool tables, flow diagrams, quick start, examples, and a distinction between standalone use and the wider ecosystem. Install instructions, Windows command differences, per-tool doctor status, FAQ, demos, and limitations are useful. The MIT license is explicit. Package version, CI, and a security policy provide some update information, but the supplied material does not include CHANGELOG contents or a clearly identified maintenance owner or support identity, limiting confidence in version history and responsibility.

The 42 structured tools, machine-readable contract, MCP interface, JSON output, probe-and-verify workflow, and platform templates can produce inspectable media deliverables. Local processing reduces the need to upload footage. The standard-library implementation and lack of API keys reduce adoption cost, while requiring users to install a compatible FFmpeg and sometimes optional components, fonts, or ASR models leaves configuration and maintenance costs.

Many claims point to specific repository material such as the package manifest, CI workflow, test fixtures, security policy, and README feature descriptions. However, the supplied evidence omits full source, CHANGELOG contents, and revision-specific CI results, limiting cross-source confirmation. The README separates capabilities from limitations and says visual judgment is performed by the calling agent while highlight ranking uses proxies. Confidence remains low for this static review.