oxbshw/watch-skill

Give AI agents eyes, ears, and verifiable results. Watch Skill turns video, audio and screen activity into searchable, timestamped evidence and proves work with deterministic contracts, not model opinion. DeepWatch is the agent workspace built on DeepSeek Harness. Python + npm, MCP, CLI, REST, Web.

What it solves

Watch Skill provides AI agents with the ability to perceive video, audio, and screen activity and verify the results of their actions through deterministic contracts. It solves the problem of agents relying on model opinions rather than verifiable evidence when interacting with multimedia sources or performing tasks.

How it works

The project consists of two primary capabilities:

  1. Perception: It transforms video, audio, and screen activity into frames, transcripts, and OCR text, each linked to an absolute timestamp. This creates a searchable index of a source that allows agents to cite specific moments as evidence.
  2. Verification: It uses a separate process to evaluate "frozen contracts" (such as file digests, JSON values, SQL results, HTTP responses, and DOM state) to provide a deterministic verdict (VERIFIED, FAILED, UNVERIFIED, or INCONCLUSIVE) that does not come from a language model.

It can be integrated into existing agents via MCP (Model Context Protocol), a CLI, or a REST API. It also includes a dedicated workspace called DeepWatch, built on the DeepSeek Harness, which integrates these tools into a ready-made agent environment.

Who it’s for

  • AI Agent Developers: Those who want to give their agents the ability to "see" and "hear" multimedia content and prove their work.
  • Users of MCP Clients: Users of tools like Claude Code, Cursor, or Zed who want to add perception and verification capabilities to their existing agents.
  • Developers fixing bugs in UI/UX: Those using "THE LOOP" (observe, critique, fix, verify) to automate the capture and verification of visual faults in web applications.

Highlights

  • Deterministic Verification: Results are verified by a separate process, not by the LLM, ensuring accuracy.
  • Timestamped Evidence: Every answer provided by the agent cites a specific timestamp in the video or audio source.
  • THE LOOP: A specialized workflow for capturing a browser session, critiquing the output via a vision model, fixing code, and verifying the fix through re-capture.
  • Broad Agent Integration: Supports over 25 agents via MCP, including Cursor, Claude Code, and VS Code.
  • Local-First: Lessons learned from corrections are stored locally and persist between runs without being uploaded.

Related

  • Project
    kangarooking/kangarooking-skills

    A collection of structured, open-source Agent Skills compatible with the Agent Skills standard, providing reusable workflows for content creation, image generation, and AI project engineering.

  • Project
    joeseesun/qiaomu-seo

    A professional SEO Agent Skill for AI assistants that performs technical audits, traffic diagnostics, and visibility optimization for traditional search engines and AI search platforms.

  • Project
    yanhua1010/self-media-content-workflow

    A modular suite of AI agent skills for self-media content production, automating the workflow from research and strategy to platform-native copywriting and multi-platform publishing.

  • Project
    agiwhitelist/auteur

    An AI agent skill that directs the creation of cinematic, high-performance websites using a disciplined pipeline of art direction, asset generation, and strict anti-slop linting.