Bottom line
Inference, medium confidence: adopt an existing web search and extraction layer, either Firecrawl self-hosted (AGPL-3.0) or a commercial API (Tavily, Exa, SerpApi) through its official SDK, and wrap it with your own quality and latency checks. The ranking machinery in this category is closed, and no repository in the evidence publishes a relevance benchmark or a latency budget, so "the search is good" is something you have to measure yourself. Keep a frozen, hashed set of recorded queries and responses as your test oracle, and keep live vendor API tests in a separate lane behind secrets rather than in the required build.
Adopt, do not rebuild
License checked on 2026-09-23; non-permissive licenses are named in the text.
- firecrawl/firecrawl AGPL-3.0: an open-source web data API to search, scrape and interact, with a self-host path and a per-commit quality-evaluation pipeline; the only open incumbent in the set with real eval plumbing. License: AGPL-3.0, so running a modified Firecrawl as a network service obliges you to publish your changes. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:7
The web data API to search, scrape, and interact at scale; fully open-source monorepo with self-host path
; ecosystem/pickup/_evidence/web-search-apis.md:37Firecrawl is the exception and the only repo in this set with a real quality-eval pipeline
; fresh read 2026-09-23,gh api repos/firecrawl/firecrawl --jq .license.spdx_id) - tavily-ai/tavily-python, exa-labs/exa-py, serpapi/serpapi-python MIT MIT MIT: official SDKs for the commercial search APIs, when a closed index is acceptable; each is a thin client you should not rewrite. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:10
Official Tavily Python SDK. Commercial vendor with open SDK
; ecosystem/pickup/_evidence/web-search-apis.md:11Official Exa Python package (neural search API)
; ecosystem/pickup/_evidence/web-search-apis.md:12Official SerpApi Python wrapper.
) - searxng/searxng AGPL-3.0: open metasearch aggregation to use as a backend under your own agent layer; it returns search-results JSON for humans and is not an agent API on its own. License: AGPL-3.0, with the same network-service source obligation for modified versions. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:8
Free metasearch engine; the open aggregation layer underneath agent search stacks
; ecosystem/pickup/_evidence/web-search-apis.md:38SearXNG is infrastructure, not an agent API.
; fresh read 2026-09-23,gh api repos/searxng/searxng --jq .license.spdx_id) - jina-ai/reader Apache-2.0: URL-to-markdown extraction endpoint with a unit and end-to-end test split; the extraction half you would otherwise write. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:9
URL→LLM-friendly-markdown endpoint
; ecosystem/pickup/_evidence/web-search-apis.md:28Unit + e2e test tiers with a booted-server e2e harness
)
Copy these practices
- A labeled quality run on every merge, against what shipped: Firecrawl triggers an evaluation against production after each deploy, labeled with the commit and an experiment ID. Copy the plumbing and write your own ground truth. Starter kit: Receipts carry provenance; generation binding. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:22
Post-deploy quality-eval benchmark against production, labeled per commit
) - On-demand evaluation from a review comment: a maintainer comment dispatches a quality eval, gated by organisation membership. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:23
Comment-triggered on-demand eval for PRs
) - The search backend as a CI matrix axis: Firecrawl's self-hosted suite already makes the search backend a matrix axis, with one value today (SearXNG; Google disabled); add a second backend so a provider swap is proven, not assumed. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:24
parameterize the search backend as a matrix axis so a clean-room implementation proves provider-swap correctness in CI
; ecosystem/pickup/_evidence/web-search-apis.md:24search backend searxng (google disabled)
) - Unit tests with HTTP interception, live tests separate: split unit and integration directories and mock transport at the HTTP layer for unit tests. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:26
new project should mock transport at the HTTP layer for unit tests
) - Every example executes in CI: Exa runs all its example scripts in CI, against live APIs with keys, so documentation cannot rot unnoticed; keep that lane out of the required build. Starter kit: Claim discipline enforced: README claims cross-checked against proofs. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:27
Docs-as-tests: every example executes in CI
; ecosystem/pickup/_evidence/web-search-apis.md:27run with EXA_API_KEY/OPENAI_API_KEY secrets against live APIs
; ecosystem/pickup/pickup-web-search-apis.md:35as oracles and never required for a green build
) - Per-option end-to-end fixture files for extraction: Jina Reader boots its server and tests content fidelity, chunking, token budgets, error handling and option interactions one file each. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:28
For any extraction endpoint, this is the file-level template.
)
Build only if
- You need control of the index and ranking: that machinery is closed in every vendor, so no incumbent offers it; but an internet-scale index is a different project type, not a web search API layer. [Verified] (ecosystem/pickup/_evidence/web-search-apis.md:36
lives inside Tavily/Exa/SerpApi/Brave/Perplexity and is not open
; ecosystem/pickup/pickup-web-search-apis.md:25Building a web index or crawler at internet scale
) - A key-free, self-hosted requirement alone does not justify building, because Firecrawl ships a self-host path. Firecrawl and SearXNG are both AGPL-3.0 (fresh read 2026-09-23,
gh api repos/<owner>/<repo> --jq .license.spdx_id); a builder who cannot accept AGPL source obligations for a modified network service still has AGPL-free adoption paths to try first: the commercial APIs through their MIT SDKs, Jina Reader (Apache-2.0) for extraction, Firecrawl's own SDKs, which are MIT, and running Firecrawl unmodified as a separate service. Only if none of those fits does the AGPL constraint justify building, and a clean-room verdict would still have to quote the clause. The other adopted repositories read MIT or Apache-2.0 in the same read. [Inference] (ecosystem/pickup/_evidence/web-search-apis.md:7fully open-source monorepo with self-host path
; stack/licenses.tsv:27SDKs under apps/-sdk (python, js, ruby, elixir) are MIT
; stack/METHOD.md:79Check the license of everything you tell a builder to adopt
)
Where FrankenSuite touches this
- frankensearch is a local two-tier hybrid search library and CLI over your own corpus, not a web search API. Its latency is receipted against its own budget, on a 1,000-document synthetic corpus on one machine, while its relevance is unproven by its own findings document, the same gap this category shows. TRL 5–6, Explore; license: MIT with a rider barring OpenAI, Anthropic and anyone acting for them. [Verified] (packets/frankensearch-assessment.md:19
A two-tier hybrid search system for Rust plus the fsfs standalone CLI
; packets/frankensearch-assessment.md:130.4 ms initial delivery vs a <15 ms target
; packets/frankensearch-assessment.md:13the receipts come from a 1,000-document synthetic corpus on one Threadripper
; packets/frankensearch-assessment.md:21Relevance quality is unproven
; packets/frankensearch-assessment.md:3MIT + OpenAI/Anthropic rider (non-OSI, rider quoted verbatim in §4.8)
; synthesis/00-overview.md:96frankensearch | 5–6 | Explore | Rider
)
What we cannot say
- How good any of these search results are: the category is mostly closed-source and no public relevance benchmark was found (ecosystem/pickup/_evidence/web-search-apis.md:36
This category is mostly closed-source.
; ecosystem/pickup/_evidence/web-search-apis.md:36No public relevance benchmark (e.g., vendor-side FreshQA-style evals, NDCG@k dashboards) was found in any repo
). Vendor benchmark percentages in the press are marketing. - How Firecrawl scores its own evals (ecosystem/pickup/_evidence/web-search-apis.md:37
the scoring methodology is not publicly documented in the repo
). - Latency: no repository asserts a latency budget (ecosystem/pickup/_evidence/web-search-apis.md:42
No latency-SLO evidence found in any repo.
). - Jina Reader's freshness (ecosystem/pickup/_evidence/web-search-apis.md:40
Jina Reader's last push (2026-05-22)
). Funding and acquisition news is press, not repo evidence (ecosystem/pickup/_evidence/web-search-apis.md:18press claims, not repo-verified
); Brave and Perplexity offer no open SDK repository (ecosystem/pickup/_evidence/web-search-apis.md:18Brave Search API and Perplexity Sonar are API-key-only products with no official open SDK repo found
). - SearXNG returns search-results JSON for humans, not agent-ready answers (ecosystem/pickup/_evidence/web-search-apis.md:38
its API returns SERP-style JSON for humans
). - Who uses these in production (ecosystem/pickup/_evidence/web-search-apis.md:44
Adopter evidence is thin at repo level.
). Pricing was not in the pack, and licenses come from the author's fresh read, not the pack; the SDK workflows were read as files, which shows they exist, not that they pass.
Revisit when
- A vendor-independent relevance benchmark for agent search is published, or Firecrawl documents its scoring method.
- Any incumbent publishes and gates a latency budget.
- Firecrawl's self-host footprint turns out to block a real deployment, or Firecrawl or SearXNG changes license.
Sources
- Pickup companion: ecosystem/pickup/pickup-web-search-apis.md, the charter seed, oracle candidates, and first claim registry for this area.
- Evidence pack: ecosystem/pickup/_evidence/web-search-apis.md, the repositories checked against the GitHub API, with dates.
- Verdict source: stack/web-search-apis.md, written under stack/METHOD.md.
Starting a build here?
Begin from the pickup companion for Web search API layers: it carries the charter seed, the oracle candidates and how to verify them, and the first claims to register, so a new project starts from evidence instead of from scratch.