Codex
Keep the key in an environment variable, add one hosted server, and verify the shared desktop, CLI, or IDE configuration.
Open Codex guidePublic beta
A YouTube research server built for agents.
Search saved research, discover channel uploads, and fetch timed transcripts, summaries, and structured insights from Codex, Claude Code, Cursor, or another Streamable HTTP client.
Install in Codex
codex mcp add jumpwatch --url https://api.jumpwatch.cc/mcp/ \
--bearer-token-env-var JUMPWATCH_API_KEYChoose the setup that matches your client.
Each guide includes the exact configuration, verification path, first prompt, and client-specific troubleshooting.
Claude Code
Install the remote HTTP server from your terminal, inspect the saved entry, and verify it inside Claude Code.
Open Claude Code guideCursor desktop
Choose global or project scope, save an authenticated mcp.json entry, and use MCP Logs when a connection fails.
Open Cursor guideGet free key
The authenticated setup also includes VS Code, Windsurf, and SDK configurations.
Small, deliberate toolset
A focused toolkit for YouTube research.
Clear verbs, bounded responses, and the cost returned on every call.
Free in beta
youtube_library_search
Search your own saved titles, channels, summaries, notes, tags, and collections with literal matching.
1 credit with results
youtube_channel_videos
List the newest official uploads for a channel handle, canonical URL, or channel ID. Empty results are free.
1 credit
youtube_transcript
Fetch a full timed transcript in text, markdown, or structured segments.
From 5 credits
youtube_summarize
Compress a video into a brief, detailed, bulleted, or chaptered summary. Cost scales with length.
5 credits
youtube_insights
Extract takeaways, topic sections, and named entities from up to 48,000 transcript characters.
Free
youtube_usage
Let an agent check its balance, reset date, limits, and tool costs before it works.
Know the boundaries before you connect.
The beta is useful because its contract is narrow and visible.
Captions are required
JumpWatch can read public creator captions or YouTube auto-captions. It cannot recover speech from a video with no caption track.
Search stays inside your library
Library search uses literal matching across your own saved metadata, summaries, notes, tags, and collections. It does not search transcript bodies, the wider YouTube catalog, or semantic similarity.
Discovery is channel-scoped
Channel discovery returns the newest official uploads for a supplied channel. Keyword discovery, playlists, and bulk ingestion are not shipped.
Long responses are bounded
Transcripts expose next_offset for paging. Structured insights analyze at most the first 48,000 transcript characters.
The meter is part of the product.
Agents retry, loop, and spend quickly. JumpWatch makes the economics visible and predictable instead of leaving you with a surprise.
Failures cost zero
Errors are returned as tool results and never consume credits.
Retries are deduplicated
An identical retry inside the safety window returns the same result without a second charge.
Balance on every result
The calling agent can see what it spent and pace the rest of its task.
Hard response limits
Long transcripts are paginated so one call cannot flood the model context.
Start with 200 free credits.
Create one key, paste one config, and make your first tool call in under three minutes.
Get free key