Skip to content

Search swarm

github-actions[bot] edited this page Nov 21, 2026 · 3 revisions

clippa/swarm.py is the search clippa story (#84) and investigate check use: many small searches at once, each free to follow up on what it finds. It is library code, not a command. A planner (Sonnet) turns the goal (a story brief, a lead) into up to 16 first tasks; workers (Haiku, 8 at a time) each run one tool in code and read its numbered results: YouTube search (bounded to the run's week, month or year, approximate dates), a YouTube channel's uploads and streams, a Kick channel's VODs, an Odysee channel, Reddit and the web (Brave), an archive subject's transcripts and our own streams' segments. Each worker marks the results that fit (strong or maybe) and proposes up to 3 follow-ups (a clip channel that keeps covering it, another person, a quote to search). The coordinator merges findings by platform id, queues follow-ups it hasn't seen as the next round, and stops at [swarm] max_rounds, max_tasks, budget_usd or max_wall_s. A channel task lists up to channel_depth videos, keeps the run's dates and passes the 20 whose titles fit the goal best. Model output never becomes a fetch: a finding is one of the tool's own results, a follow-up must parse as its kind (a query, a supported platform's channel, an archive on file), and targets asking for anyone's whereabouts or private details are dropped. The trail is the caller's swarm.json (every task, its parent, results, findings, cost); a rerun resumes it. A 30-task run on Jer's Huck story found 109 results (44 strong) across Kick, YouTube clip channels, Odysee, Reddit and the web for ~$0.03.

Setup

The swarm needs ANTHROPIC_API_KEY and yt-dlp (with deno, as for pre-shows). Its web and Reddit searches use the same Brave key; Kick, Odysee and YouTube listings need no key. YouTube searches run from this machine's IP, so a bot check here affects them too. Per-source request rates are [swarm] rates.

Clone this wiki locally