Agent compatibility

Watch Skill exposes one engine through four connection methods. Choose the method your agent supports; the underlying index, tools, and privacy settings are identical.

The agents

Twenty-six agents with a page of their own, plus frameworks and HTTP clients below. Each avatar links to that agent's setup and its verification status; the tables after this say which method it uses and how far the status has been proven.

What the status labels mean

Status	Evidence required
Machine-tested	A real agent session connected, called tools, and received results end to end.
Machine-configured	Setup wrote a valid configuration and the exact server command completed an MCP initialize handshake. An in-app agent run has not yet been recorded.
Documentation-verified	The configuration matches the agent's published documentation and every fenced config block passes the repository validator. It has not been executed here.

These labels describe integration evidence, not product quality. Run python scripts/validate_agent_docs.py to check the examples in this directory.

MCP clients

MCP is the broadest integration path. Local clients normally start watch-skill serve over stdio; remote clients use the streamable HTTP endpoint.

Agent	Configuration	Status
Claude Desktop	claude_desktop_config.json	Machine-configured
Cursor	~/.cursor/mcp.json	Machine-configured
Codex CLI	~/.codex/config.toml	Machine-configured
Cline	MCP settings UI / cline_mcp_settings.json	Documentation-verified
Windsurf	~/.codeium/windsurf/mcp_config.json	Documentation-verified
Gemini CLI	~/.gemini/settings.json	Documentation-verified
VS Code Copilot	.vscode/mcp.json	Documentation-verified
GitHub Copilot CLI	~/.copilot/mcp-config.json	Documentation-verified
Kimi Code CLI	~/.kimi-code/mcp.json or kimi mcp add	Documentation-verified
Qwen Code	~/.qwen/settings.json	Documentation-verified
OpenCode	opencode.json	Documentation-verified
Goose	~/.config/goose/config.yaml	Documentation-verified
OpenHands	config.toml	Documentation-verified
Kilo Code	kilo.jsonc	Documentation-verified
Qodo Command	Project mcp.json and agent TOML	Documentation-verified
Agent Zero	Settings UI, stdio or HTTP	Documentation-verified
Zed	settings.json — context_servers, not mcpServers	Documentation-verified
Roo Code	.roo/mcp.json or global mcp_settings.json	Documentation-verified
Continue	.continue/mcpServers/watch-skill.json	Documentation-verified
JetBrains IDEs	.junie/mcp/mcp.json, or AI Assistant settings	Documentation-verified
Amp	amp mcp add, or the amp.mcpServers settings key	Documentation-verified
Aider	No MCP client — runs the CLI through /run	Documentation-verified

Plugin and skill-native agents

Skill-native integrations add trigger guidance as well as tools. They teach the agent when to watch, when to ask the existing index, how to cite evidence, and when to verify its own work.

Agent	Integration	Status
Claude Code	Plugin with ten skills and MCP	Machine-tested
OpenClaw	SKILL.md discovery	Documentation-verified
Pi	Skills directory and CLI	Documentation-verified
Hermes Agent and similar harnesses	Skills, AGENTS.md, or REST	Documentation-verified
Any instruction-following coding agent	AGENTS.example.md	Machine-tested in this repository

Frameworks and HTTP clients

Python adapters wrap watch_video, ask_video, and search_videos as native framework tools. TypeScript and automation platforms use the REST/OpenAPI surface. The complete setup and code samples are in the framework adapter guide.

Framework or client	Integration	Status
LangChain / LangGraph	Native Python tools	Machine-tested
CrewAI	Native Python tools	Machine-tested
OpenAI Agents SDK	Native Python tools	Machine-tested
LlamaIndex	Native Python tools	Unit-tested
AutoGen 0.4+	Native Python tools	Unit-tested
Vercel AI SDK	REST tool	Documentation-verified
n8n	HTTP Request node / webhook	Documentation-verified
Any HTTP client	REST + OpenAPI	Machine-tested
Any remote MCP client	Streamable HTTP at /mcp	Machine-tested

Fast path

The installer runs watch-skill setup, which detects supported agents and offers to write their configuration. Existing files are backed up before modification.

watch-skill setup
watch-skill doctor

After restarting the agent, use the same smoke test everywhere:

1. Confirm that watch-skill appears in the client's tool or MCP list.
2. Ask it to watch a short public video and describe a specific timestamp.
3. Ask a follow-up about the same video. The agent should use ask_video, not process the source again.

Each linked guide supplies the exact configuration and client-specific verification steps.

Add another agent

Start with the agent integration template. A complete row needs one working config block, one three-step smoke test, and an honest status label. See CONTRIBUTING.md for the review checklist.