Firecrawl released Developer Index on August 20, a search layer built specifically for coding agents. It indexes more than 70 million artifacts, including READMEs, issues, pull requests and external documentation, and answers natural-language questions with the matching passages attached. Alongside it, the company published DevDex, an open benchmark for measuring developer-search quality.

What agents are looking for is not a web page

Coding agents burn a large share of their tool calls on lookups. Three patterns dominate: the repository that implements an idea, the documentation page that answers a question, and the issue or pull request where a bug was discussed and fixed.

The trouble is that all three live in different places, scattered across GitHub, documentation sites and threads. Existing search is largely lexical, so it misses things. And pulling a README together with its issues and recent pull requests can mean stitching together more than 50 API calls, Firecrawl says.

The company also describes the customers it built this for: agentic products that debug backends on behalf of end users, teams stitching internal and external repositories into a single retrieval layer, and frontier labs that need open developer documentation, code, issues and pull requests as training and evaluation data. All of them were working around the same gap.

It indexes artifacts, not code

Developer Index holds more than 70 million artifacts: READMEs, pull requests, issues, OpenAPI specs, skills and external documentation. Refreshes run continuously, and most sources are updated daily.

Issues and pull requests come from top repositories along with their linked artifacts. READMEs cover a broad set of public repositories, and external documentation includes well-structured sources of the kind Stripe publishes. Every artifact carries metadata such as star count, license and artifact type.

The index does not store code, and it is not a general web search endpoint. Firecrawl draws the line clearly: this is a retrieval layer organized around the artifacts coding agents produce and consume.

Two entry points, and no API key needed to start

Usage is straightforward. Send a natural-language question and you get ranked developer results with the passages that matched. Passages come back as Markdown, so tables and code blocks survive intact and the agent can act without a second scrape.

There are two ways in: a dedicated endpoint that returns developer sources only, and the standard search endpoint with a developer category specified. Each result carries a stable ID whose prefix tells you whether it is a document, an issue, a pull request or a README.

Through the API you can filter by type, repository, source, language, topic, license and minimum star count, and there is a setting that searches indexed agent-skill files only. Those filters are API-only. On the CLI and MCP surfaces, agents actually perform better without them, so Firecrawl deliberately leaves them out.

No API key is required to try it, and adding one raises the rate limits. Pricing is 2 credits per 10 results, rounded up.

npx -y firecrawl-cli@latest setup developer-index

It ships through the API, CLI, MCP and the SDKs, and plugs into harnesses already in use, including Codex, Claude Code and Grok Build.

A yardstick shipped alongside the product

The more interesting move is that Firecrawl published a benchmark at the same time. DevDex exists because standard search benchmarks do not reflect what agents actually look up while writing code.

It consists of 1,179 developer-search queries across three tracks: repository discovery, documentation lookup, and issue and pull request resolution. Scoring uses Recall@10, which asks whether the correct artifact appears anywhere in the top ten, and MRR@10, which accounts for rank. Grading is deterministic against fixed references, and a memorization check drops any query the driver model can answer from pretraining alone. Every provider runs under a matched setup with Claude Opus 4.8 as the driver model and one search tool active per run.

The results: Developer Index scores 0.63 overall. Firecrawl Search with no category follows at 0.58, Parallel at 0.57, and Mintlify and Exa at 0.54. Native web search lands at 0.45 and Context7 at 0.17.

Broken out by track, the strengths and weaknesses show. Developer Index leads issue and pull request resolution at 0.66 and is effectively tied with Context7 on documentation lookup at 0.47. Repository discovery is its weakest track at 0.76, behind Parallel at 0.82 and Firecrawl Search at 0.78. Context7, which is docs-focused, scores 0.01 on repository discovery and 0.03 on issues and pull requests, a narrow range that shows up directly in the numbers.

Firecrawl has open-sourced half the dataset and the evaluation harness so other teams can measure their own systems on the same ground.

Summary

Developer Index is an attempt to give coding agents their raw material as artifacts rather than as web pages. An index of more than 70 million items, refreshed daily and returned with matched passages, is aimed squarely at cutting the number of round trips an agent needs. It takes the overall lead in the benchmark while ceding repository discovery to rivals, so the right choice still depends on the workload. Publishing the benchmark alongside the product suggests developer search is finally becoming something teams can compare.