firecrawl-developer-search

firecrawl developer search

Via the Firecrawl MCP (preferred)

This plugin bundles the Firecrawl MCP server. When it's connected, prefer the native firecrawl_developer_search tool over the CLI — it needs no local install and authenticates through the plugin's one-time browser sign-in. The firecrawl CLI commands below are the fallback for when the MCP isn't connected.

Semantic search over indexed developer sources. Each result carries an id, a url, a title, and the matched passages in markdown, so tables and code blocks survive intact.

The id prefix tells you what kind of source a hit is: doc:, issue:, pull_request:, or readme:.

When to use

The question is about code behaviour, a library or framework, an API contract, an error message, or a known bug
You want the issue that reported it, the merged PR that fixed it, or the doc page that defines the contract — not a blog post describing it
Before falling back to a general web search: reach for the index first, the open web second

Not for this skill: comparisons and opinions ("X vs Y"), news and outages, or a project with no public repo or indexed docs. Those are firecrawl-search plus firecrawl-scrape questions.

Quick start

MCP (preferred):

firecrawl_developer_search(query: "tokio select cancellation safety", k: 10)
firecrawl_developer_search(query: "agent skill frontmatter conventions", skills: "only")

CLI fallback:

# Ask a developer question
firecrawl developer "axum middleware ordering" --limit 10 -o .firecrawl/dev-axum-middleware.json --json

# Search a literal error string plus the library name
firecrawl developer "ERR_REQUIRE_ESM in vitest config" -o .firecrawl/dev-esm-error.json --json

Put all scoping intent in the query text — the repo slug, the library name, the version, the error string. Semantic retrieval handles the scoping; there are no repo or type flags on the MCP tool or the CLI.

Options

Option (MCP)	Option (CLI)	Description
`query`	`<query>`	Natural-language question or search phrase (required)
`k`	`--limit <n>`	Number of ranked results, 1–100 (default 10)
`skills`	—	Set to `"only"` to search only agent-skill files
—	`-o, --output <path>`	Output file path
—	`--json` / `--pretty`	JSON output, compact or pretty

Tips

Match the query to the question. A literal error message wants the error string plus the library name. A conceptual "how do I do X" wants the whole question in natural language.
Volatile parts hurt a match. If an error string returns nothing, strip paths, line numbers, ids, and addresses and retry — the invariant middle of the message is what's indexed.
A merge supersedes a report. When an issue and a pull request disagree, the merged PR is the current behaviour. Read the resolution before answering, never the opening report alone.
Quote the passage, cite the url. The passages are the evidence — hand them over rather than paraphrasing into a claim the user can't check. title is often absent on doc: hits; fall back to url.
Raise k before giving up, and scrape when a hit is the right page. If a result is clearly the right page but the passage is the wrong part of it, scrape its url for the whole thing.
Already running a web search? firecrawl_search(query, categories: ["developer"]) folds developer hits into the same call as web results (they come back tagged category: "developer"). It exposes no passage control, so prefer the dedicated tool when the question is purely a code question.
With the CLI, write results to .firecrawl/ with -o to keep the context window clean. Naming convention: .firecrawl/dev-{topic}.json.

About this skill