web-search-tool 0.1.0

Unified CLI over Tavily, Firecrawl, and SerpAPI web-search/content SDKs

web-search-tool

A unified Typer CLI over the Tavily, Firecrawl, and SerpAPI web-search/content SDKs — one command surface and one normalized output schema for humans, scripts, and agents.

• One schema, every provider. search and every vertical return the same envelope shape (title, url, snippet, score) plus documented vertical-specific fields, regardless of which provider served the request.
• A resilient fetch cascade. fetch walks markdownify → tavily → firecrawl, falling through on thin content, hard HTTP errors, and unsupported sites.
• Agent-friendly. Deterministic exit codes, JSON when piped, an explicit escalation signal on total fetch failure, and byte-bounded/chunked output.

Install

cd web-search-tool
uv sync --extra dev
uv run web-search-tool --version

Configuration

Settings resolve with the precedence flag > env > config file > default.

• Tool settings use the SEARCH_TOOL_ env prefix (e.g. SEARCH_TOOL_DEFAULT_PROVIDER=tavily, SEARCH_TOOL_TIMEOUT=30).
• Provider keys use their conventional names: TAVILY_API_KEY, FIRECRAWL_API_KEY, SERPAPI_API_KEY.
• Set a key to SKIP to opt a provider out of launch validation. Requesting a vertical that only that provider serves (e.g. scholar/jobs with SERPAPI_API_KEY=SKIP) fails with a specific SKIP-conflict error.
• An optional TOML config lives at ~/.config/web-search-tool/config.toml (override with WEB_SEARCH_TOOL_CONFIG or the --config flag). See config.example.toml for every tunable.

# Minimal: keys in the environment
export SERPAPI_API_KEY=...        # required for the SerpAPI verticals
export TAVILY_API_KEY=...         # search, research, one cascade converter
export FIRECRAWL_API_KEY=SKIP     # opt out if you don't have a Firecrawl key

Setting	Env var	Default	Purpose
Default provider	SEARCH_TOOL_DEFAULT_PROVIDER	serpapi	Provider for generic search/resolve
Timeout	SEARCH_TOOL_TIMEOUT	20	Per-request timeout (seconds)
Default format	SEARCH_TOOL_DEFAULT_FORMAT	(auto)	table or json; auto-detects TTY when unset
Tavily key	TAVILY_API_KEY	—	Tavily search + deep research
Firecrawl key	FIRECRAWL_API_KEY	—	Firecrawl search + scrape converter
SerpAPI key	SERPAPI_API_KEY	—	All seven SerpAPI verticals
Sentry DSN	SENTRY_DSN	(off)	Enables error tracking + tracing when set
Sentry environment	SENTRY_ENVIRONMENT	local	Deployment tag on Sentry events
Trace sample rate	SENTRY_TRACES_SAMPLE_RATE	1.0	Transaction sampling, 0.0–1.0

The Sentry settings also live under a [monitoring] table in the config file and follow the same precedence. See Monitoring.

Output contract

• stdout carries only the payload. All logs go to stderr; API keys never appear in logs, output, or written files.

• Format: a TTY renders a rich table; a pipe emits JSON. --format/-f table|json overrides both; --compact removes JSON indentation.

• Every payload is wrapped in an envelope — {command, ok, data} — so callers can branch on ok without parsing the body.

• Exit codes are deterministic and category-specific:

  Code	Name	Meaning
  0	SUCCESS	Completed
  1	GENERAL	Unclassified error
  2	USAGE	Bad CLI usage
  3	INPUT	Bad input / no URL resolved
  4	NOT_FOUND	No results
  5	NETWORK	Network/provider failure
  6	TIMEOUT	Bounded wait exceeded
  7	CONFIG	Missing/SKIP key, bad config

Commands

Search & verticals

web-search-tool search "python typer cli" --limit 10 [--provider tavily|firecrawl|serpapi]
web-search-tool news "us treasury yields" -n 5
web-search-tool jobs "platform engineer" --location "Boston, MA"
web-search-tool images "golden gate bridge"
web-search-tool videos "rust async tutorial"
web-search-tool reverse-image "https://example.com/photo.jpg"
web-search-tool scholar "continual learning llm" \
    --author "Bengio" --min-cites 50 --since 2020 --until 2024 --sort cites

All search commands accept --limit/-n, --format/-f, and --compact. --limit is enforced uniformly — SerpAPI verticals are capped client-side.

Scholar filters (scholar only):

Flag	Effect
--author/-a NAME	Native author:"NAME" operator
--min-cites N	Keep only results cited ≥ N times (client-side over a pool)
--since YEAR / --until YEAR	Native publication-year range
--sort relevance|date|cites	relevance (default), most-recent, or most-cited

Fetch — URL → markdown via the cascade

web-search-tool fetch "https://example.com/article"        # probe: first success wins
web-search-tool fetch URL --compare                        # all converters, labeled
web-search-tool fetch URL --raw-html                        # local markdownify only
web-search-tool fetch URL --order markdownify,firecrawl     # custom chain
web-search-tool fetch URL --max-bytes 20000 --offset 0      # bounded / chunked window
web-search-tool fetch URL --stdout                          # stream instead of writing a file

• Default mode is a probe: walk the chain, first success wins; total failure raises chain-exhaustion carrying the exact signal Fetch chain completely failed, try using agent-browser.
• --max-bytes overflow truncates and appends the marker *OUTPUT TRUNCATED PLS INCREASE THE CAP OR DO CHUNKED REQUEST*; combine with --offset for deterministic chunked continuation.
• Without --stdout, output is written to a collision-safe slug file (<UTC-timestamp>-<slug>.md) in --out-dir (default: current directory).

Resolve — search, then fetch the top-N URLs

web-search-tool resolve "best rust web framework" --limit 5
web-search-tool resolve "fed rate decision" --vertical news -n 3
web-search-tool resolve "q" --stdout

Runs a search over --vertical (default search), then fetches each result URL through the cascade in auto-resolve mode: a URL that exhausts its chain records an escalation but never aborts the batch. Exits 0 if any URL resolved, 3 (INPUT) only if all failed.

Research — deep cited research via Tavily

web-search-tool research run "history of CRISPR patents"            # returns request_id
web-search-tool research run "q" --file notes.md --file data.json   # attach local context
web-search-tool research run "q" --wait --max-wait 300              # poll to completion
web-search-tool research status <request_id>                        # look up once

research run returns a request_id immediately; --wait polls until a terminal status or --max-wait (then exits TIMEOUT with a resume hint). research status reports status and, on completion, the cited markdown report with numbered sources. --file attaches up to 5 local .txt/.md/.json files as research context.

Monitoring

Sentry is off by default and fully optional — set a DSN to turn it on. When enabled, the tool reports errors and wraps each invocation in a transaction with child spans around every outbound call (provider search, each fetch-cascade tier, research launch/poll), tagged with non-secret data (provider, tier, status, result counts). API key values are never attached to spans, tags, or events.

Configure it via env or the [monitoring] table in the config file (same precedence as everything else):

export SENTRY_DSN="https://<key>@<org>.ingest.sentry.io/<project>"
export SENTRY_ENVIRONMENT=prod          # optional, defaults to "local"
export SENTRY_TRACES_SAMPLE_RATE=0.2    # optional, defaults to 1.0

Verify the setup end to end — emits a span tree plus one synthetic, harmless exception, then prints the captured event id:

web-search-tool test sentry

It exits with the config code (7) when no DSN is configured (nothing to test).

Development

uv run pytest                          # unit tests (live tests skipped without keys)
uv run pytest -m live                  # opt-in tests against real provider APIs
uv run ruff check . && uv run ruff format --check .
uv run mypy --strict src/web_search_tool
uv run pytest --cov=web_search_tool --cov-report=term-missing

See CONTRIBUTING.md for the full workflow.

License

MIT © Vlad Korolev