Skip to content

Command line

swarm research QUESTION

Flag Default Meaning
--preset quick\|standard\|deep standard breadth, depth and default cost cap ($0.75 / $2 / $6)
--search tinyfish\|anthropic tinyfish free TinyFish tools, or Claude's server-side search ($10 per 1,000 searches)
--docs PATH... your files or folders as sources (txt, md, html, pdf)
--docs-only no web: research only --docs
--include-domains / --exclude-domains excludes pinterest.com, quora.com comma-separated domains
--recency-days N only recent results
--news / --papers search news or research papers
--language CODE language for searches and the report
--out FILE stdout .md, .html or .json (.html/.md also write a .json for follow-ups)
--follow-up REPORT.json extend an earlier report
--max-usd X preset hard dollar cap (0 = none)
--estimate print the estimate and exit
--model / --worker-model claude-opus-5-5 models for planning and writing, and for workers
--provider anthropic\|openai anthropic Claude, or any OpenAI-compatible API (see Configuration)
--base-url URL $OPENAI_BASE_URL OpenAI-compatible endpoint, e.g. https://host/v1 (implies --provider openai)
--api-key-env NAME OPENAI_API_KEY environment variable that holds the provider key, e.g. RIVER_API_KEY
--price IN,OUT[,CACHED] USD per million tokens, for a model with no built-in price
--max-subquestions, --max-agents, --max-searches preset fine-grained limits
--no-claim-check skip the claim check (cheaper, less trustworthy)
--no-cache bypass the local search/fetch cache
--mock offline demo
-v print full run statistics

Other commands

Command What it does
swarm serve [--host --port --no-worker] web UI and REST API, with one embedded worker unless --no-worker; refuses a public interface without a key
swarm worker [--concurrency N] runs queued server jobs; run several beside swarm serve --no-worker
swarm keys create\|list\|revoke server API keys: --name, --daily-usd, --rpm, --admin
swarm doctor [--online] check install, keys and config
swarm eval [questions.jsonl] known-answer evaluation against a single-agent baseline (--estimate, --mock, --max-usd-total)
swarm bench offline benchmark of the scheduling engine
swarm run tasks.jsonl many small tasks: text, code checked by tests, or web lookups with cited sources (--allow-code-execution, --max-usd, provider flags as above); see Batch tasks