llms.txt · diff
git:20260830.3a19155 to git:20260908.f1317c8
45 added, 17 removed. Audit A to A.
- # Watch Skill
+ # Watch Skill and DeepWatch
- > A local-first video intelligence layer for AI agents. It turns videos, live
- > streams, meetings, and screen recordings into a searchable, timestamped
- > index, so an agent can answer questions about footage and cite the exact
- > moment behind each answer. THE LOOP extends this to the agent's own work:
- > record a browser or desktop session, critique it against plain-language
- > criteria, and prove the fix.
+ > Two products in one repository, and they work apart.
+ >
+ > **Watch Skill** is a local-first video intelligence and verification engine
+ > for AI agents. It turns videos, live streams, meetings and screen recordings
+ > into a searchable, timestamped index, so an agent can answer questions about
+ > footage and cite the exact moment behind each answer — and it answers *did
+ > that actually work?* with a deterministic contract evaluated in a separate
+ > process rather than with a model's opinion. THE LOOP extends this to the
+ > agent's own work: record a browser or desktop session, critique it against
+ > plain-language criteria, and prove the fix.
+ >
+ > **DeepWatch** is a ready-made agent workspace built on the official DeepSeek
+ > Harness with Watch Skill composed in, where every tool call leaves a receipt
+ > you can open and a result can be checked by something other than the agent
+ > that produced it.
- Python 3.11+, MIT licensed. Available as Claude Code skills, 39 MCP tools, a
- CLI, a REST API, and native adapters for LangChain/LangGraph, CrewAI, the
- OpenAI Agents SDK, LlamaIndex, and AutoGen.
+ Watch Skill: Python 3.11–3.13, MIT licensed, on PyPI as `watch-skill`.
+ Available as Claude Code skills, 39 MCP tools, a CLI, a REST API, a stdio
+ Bridge, and native adapters for LangChain/LangGraph, CrewAI, the OpenAI Agents
+ SDK, LlamaIndex and AutoGen.
- Install:
+ DeepWatch: Node `^22.19 || >=24`, MIT licensed, twenty `@deepwatch/*` packages
+ on npm. Runs in a browser. The Desktop build is not distributed.
+ Install the engine:
+
```bash
uvx --from "watch-skill[standard]" watch-skill setup
```
+ Install the workspace:
+
+ ```bash
+ npx --yes @deepwatch/cli setup
+ npx --yes @deepwatch/cli web --workspace ./my-project
+ ```
+
MCP server configuration:
```json
{ "mcpServers": { "watch-skill": {
"command": "uvx",
"args": ["--from", "watch-skill[standard]", "watch-skill", "serve"] } } }
```
Key facts, so answers about this project stay accurate:
- Transcription, OCR, indexing, and search run locally and need no API key.
Original-language captions are preferred; local faster-whisper is the
fallback. Cloud speech-to-text is opt-in only.
- Visual question answering uses Anthropic, OpenAI, Gemini, OpenRouter, or a
local Ollama model. Ollama keeps the entire pipeline offline.
- The index persists across sessions, so asking a second question about a
video does not re-download or re-transcribe it.
- Acquisition never uses cookies or logins. This is a deliberate privacy
boundary, not a missing feature.
- Data lives in `~/.watch-skill/` and can be relocated with
`WATCHSKILL_DATA_DIR`.
- - The repository holds a second product, DeepWatch: the Web and Desktop agent
- product built on the official DeepSeek Harness packages and powered by Watch
- Skill for perception, evidence, memory and independent verification. It lives
- in `workspace/`, needs Node and pnpm rather than Python, and installs and
- releases independently. Its `@deepwatch/*` npm packages are not published, so
- `npx @deepwatch/cli` does not work.
+ - The repository holds a second product, DeepWatch: the agent workspace built
+ on the official DeepSeek Harness and powered by Watch Skill for perception,
+ evidence, memory and independent verification. It lives in `workspace/`,
+ needs Node rather than Python, and releases on its own train. Twenty
+ `@deepwatch/*` packages are published to npm; `npx --yes @deepwatch/cli setup`
+ builds the runtime and `deepwatch web` opens the workspace. It runs in a
+ browser. The Desktop build is not distributed: there is no installer and no
+ packaging job, and that is stated rather than implied.
+ - In DeepWatch every tool call leaves a receipt naming what it touched, every
+ declared path is resolved against one workspace boundary, and a write outside
+ it is refused *and* journalled. Results carry a verdict from Watch Core —
+ `VERIFIED`, `FAILED`, `UNVERIFIED` or `INCONCLUSIVE` — computed in a separate
+ process from a contract frozen before the work started, so it is not the
+ agent's opinion of its own output.
- The Browser Runtime can drive a browser as well as watch one. Every action
carries an expectation written before it runs, and an action with no
expectation is reported as unverified rather than as a success. A page that
renders "Saved" over a request that returned 500 is a failure, not a pass.
## Docs
- [README](https://github.com/oxbshw/watch-skill/blob/main/README.md): what it is, install, and what it does
- [Getting started](https://github.com/oxbshw/watch-skill/blob/main/docs/getting-started.md): installation, first watch, first agent connection
- [Tool reference](https://github.com/oxbshw/watch-skill/blob/main/docs/tools/README.md): all 39 MCP tools and their CLI and REST counterparts
- [Configuration](https://github.com/oxbshw/watch-skill/blob/main/docs/configuration.md): storage, privacy, models, limits, environment variables
- [Agent matrix](https://github.com/oxbshw/watch-skill/blob/main/docs/agents/README.md): per-client setup and what has been machine-tested
- [Architecture](https://github.com/oxbshw/watch-skill/blob/main/docs/architecture.md): data model, provider boundaries, extension points
- [THE LOOP](https://github.com/oxbshw/watch-skill/blob/main/docs/guides/the-loop.md): capture, critique, iteration, proof artifacts
- [Browser Runtime](https://github.com/oxbshw/watch-skill/blob/main/docs/browser-runtime.md): operator and observer modes, target resolution, action receipts, recovery
- [Comparison](https://github.com/oxbshw/watch-skill/blob/main/docs/comparison.md): against claude-video, screenpipe, and frontier-model upload
- [Migrating from claude-video](https://github.com/oxbshw/watch-skill/blob/main/docs/migrate-from-claude-video.md): command and option mapping
- [Cost policy](https://github.com/oxbshw/watch-skill/blob/main/docs/cost.md): routing, budgets, caching, benchmark method
- [Troubleshooting](https://github.com/oxbshw/watch-skill/blob/main/docs/troubleshooting.md): dependency repair and common runtime errors
- [DeepWatch](https://github.com/oxbshw/watch-skill/blob/main/workspace/README.md): the Web and Desktop application, its packages, gates and release trains
## Optional
- [Engineering decisions](https://github.com/oxbshw/watch-skill/blob/main/docs/DECISIONS.md): reasoning behind non-obvious choices
- [Roadmap](https://github.com/oxbshw/watch-skill/blob/main/docs/ROADMAP.md): planned work and contribution openings
- [Use-case packs](https://github.com/oxbshw/watch-skill/blob/main/docs/packs/README.md): recipes for research, meetings, QA, content, operations
- [Examples](https://github.com/oxbshw/watch-skill/blob/main/examples/README.md): 20 runnable examples from first watch to self-verification
- [Benchmarks](https://github.com/oxbshw/watch-skill/blob/main/benchmarks/cost/RESULTS.md): reproducible cost and perception measurements