Watch Skill · DeepWatch
An MCP skill that turns video, audio and screen recordings into searchable, timestamped evidence for AI agents, and checks a frozen contract of file digests, JSON values, SQL or HTTP results by a separate process rather than by model opinion.
Watch Skill has two separable capabilities. Perception indexes a recording once into timestamped frames, transcripts and OCR text, so an agent can query it for as long as it is kept and every answer cites a moment that can be reopened. Verification evaluates a frozen contract, file digests, JSON values, SQL results, HTTP responses or DOM state, through a separate process rather than a language model, returning one of four verdicts: VERIFIED, FAILED, UNVERIFIED or INCONCLUSIVE.
DeepWatch is a ready-made agent workspace built on the official DeepSeek Harness with Watch Skill already composed in; every tool call leaves an openable receipt, every declared file path is checked against a single workspace boundary so a tool cannot write outside it, results carry a verdict the Library keeps across restarts, and a Compare view puts two runs of the same contract side by side. Either capability is usable on its own, reachable over MCP, a CLI or a REST API, and distributed as both a Python (PyPI) and Node (npm) package.