Firecrawl Developer Index: agents search code, not Google

Your coding agent is only as good as what it can find. Firecrawl just shipped Developer Index, a search API built for coding agents that indexes over 70 million primary developer sources — repository READMEs, issues, merged pull requests, and curated documentation sites — instead of sending your agent out to general web search and hoping it lands on a decent Stack Overflow answer. It is the least glamorous fix in agentic coding and probably the highest-leverage one.

What actually happened

Per Firecrawl's product page, Developer Index covers 70M+ docs, READMEs, issues, and merged pull requests, with daily refresh on most sources. Access is a GET or POST to /search/developer, or POST /search with categories: ["developer"] if you want developer results mixed into standard web results. There is a CLI (firecrawl developer [query]), Python and Node SDKs, and an MCP server.

The filters are where it earns its keep. The docs list types to restrict to doc, issue, pull_request, or readme; repos and sources to scope to specific projects; and repository attributes including language, topic, license, min_stars, max_stars, archived, and fork. Results carry stable type-prefixed IDs and matched passages in markdown. Pricing is 2 credits per 10 results, rounded up. You can query without an API key; a key raises your rate limit.

On Firecrawl's own DevDex benchmark — 1,179 real developer queries across repository discovery, issues and PRs, and documentation lookup — the index posts 0.63 recall@10 against 0.58 for Firecrawl's own general search, 0.57 for Parallel, and 0.54 each for Mintlify and Exa. Worth naming the obvious: this is a vendor-run benchmark on a vendor-defined query set. The margin over the next provider is roughly six points, not an order of magnitude.

Why retrieval matters for your business

Bad context costs more than bad models. When an agent hallucinates an API signature, the usual reflex is to reach for a bigger model. The cheaper fix is usually feeding it the actual README and the merged PR that changed the signature. A model reading the right page beats a smarter model guessing.

archived and min_stars are the real feature. Half of what wrecks agent output is confidently-cited dead code — a 2019 fork with eleven stars presented as canonical. Filtering by repository health at query time is a guardrail you cannot get from a general web search.

Merged PRs answer questions docs never will. "Why does this behave differently in v3" is answered in a diff and a review thread, not in a docs page someone forgot to update. Indexing merged PRs is the interesting part of this release.

Do not hardwire one index. Put retrieval behind your own interface. Firecrawl leads a benchmark this month; someone else leads it next month, and you should be able to swap without touching agent logic.

Sources: Firecrawl Developer Index, Firecrawl docs.