Firecrawl co-founder Eric Ciarla announced on August 28, 2026 that Firecrawl has relaunched as "free keyless" — AI agents can search and scrape the web "100% free," with no API key and no signup. The pitch bundles three claims: state-of-the-art search accuracy at 94.7% on SimpleQA, sub-3-second scrapes that turn any page into clean Markdown, and zero onboarding friction. One command scaffolds it:
npx -y firecrawl-cli@latest init --all --browser
It is billed as a relaunch because Firecrawl was keyless in its earliest days before it added accounts and quotas. The move also sharpens its rivalry with Exa, whose pitch is embeddings-based search; Firecrawl's founder claims the edge is accuracy plus the fact that it returns full page content, not just links. This post looks at what "keyless" actually means for a builder, how it fits a Claude Code or agent-harness setup, and the limits the announcement glosses over.
TL;DR — the questions people are asking
| Question | Short answer |
|---|---|
| Do I need an API key or account? | No. init scaffolds config; search and scrape work unauthenticated. |
| Is it really unlimited? | No — keyless removes signup, not per-IP rate limits or fair-use caps. |
| Is 94.7% SimpleQA meaningful? | It's a narrow fact-lookup benchmark. Good signal, not a full picture. |
| How is it different from Exa (besides cost)? | Firecrawl finds and returns cleaned full content; Exa is search/retrieval. Other tradeoffs need your-query testing. |
| Does it bypass robots.txt / site ToS? | No. Those still apply to what you fetch, keyless or not. |
| How does it plug into agents? | Via the Firecrawl MCP server, the CLI, or direct HTTP calls in a code harness. |
| Safe for production? | Prototype on keyless; move to authenticated or self-hosted before anything customer-facing. |
What "keyless" actually means
"No API key, no signup" is a real friction reduction — you skip account creation, secret provisioning, and key rotation. For a demo, a one-off research script, or a workshop exercise, that is the difference between "running in 30 seconds" and "running after you find the dashboard."
What it does not mean:
- No rate limits. A free, unauthenticated endpoint has to throttle per IP or it becomes a free proxy for scrapers worldwide. Expect a request-per-minute ceiling and possibly a daily cap. The announcement does not publish the numbers, so measure them before you design around them.
- No terms of service. Firecrawl's own usage terms still bind you, and — more importantly — the target sites' terms and
robots.txtstill govern what you are allowed to fetch and store. Keyless changes who Firecrawl bills, not what the wider web permits. This is the same tension covered in publishers moving to block AI crawlers. - No reliability guarantee. There is no contract behind a free tier. Limits can tighten, auth can come back, and a heavy-traffic day can degrade latency. That is a fine tradeoff for prototyping and a bad one for a shipped product.
So the accurate framing: keyless is a generous evaluation tier with the sign-up wall removed, not a production SLA.
The SimpleQA number, in context
94.7% on SimpleQA is a strong score, but SimpleQA is a deliberately narrow benchmark: short factual questions, each with a single verifiable answer ("What year did X happen?"). A high score means the retrieval layer is good at fact lookup. It says little about:
- Multi-hop research — questions that require chaining several searches and reconciling sources.
- Freshness — how quickly new pages enter the index.
- Long-tail domains — niche forums, docs sites, non-English content.
- Breadth tasks — filling a table with 40 rows, where recall across many sources matters more than pinpoint accuracy.
OpenRouter's cross-provider web-search benchmarks found that the biggest lever on agent search quality was not the engine — it was the search budget, meaning how many retrieval turns the agent was allowed. A keyless tier with an aggressive rate limit effectively caps that budget, which may matter more to your results than the headline accuracy figure.
Firecrawl vs Exa — the honest comparison
The founder's stated differentiator is accuracy plus the scrape step: Firecrawl returns the full cleaned content of the pages it finds, formatted as Markdown, in one call. Exa is primarily a neural/embeddings search API optimized for finding relevant URLs and snippets, with content retrieval as a secondary feature.
Beyond cost and that architectural difference, the tradeoffs are genuinely open and worth testing yourself:
| Dimension | What to check |
|---|---|
| Index freshness | Query a page published today; see if either returns it. |
| Result diversity | Does one over-index on SEO-optimized content vs primary sources? |
| Latency at volume | Sub-3s single scrape is not the same as 200 concurrent scrapes. |
| Quota behavior | What happens at the rate limit — 429, queue, or silent truncation? |
| Structured extraction | Both offer "describe what you want" modes; accuracy varies by schema. |
Run a real subset of your own queries through both. A single benchmark, especially a narrow one, is not a substitute. For the broader shift toward programmable, agent-first search, see Perplexity's Search as Code.
How it slots into a Claude Code / MCP / harness setup
There are three integration paths, in rough order of how most agents will use it:
- MCP server. Wire the Firecrawl MCP server into your MCP host (Claude Code, Claude Desktop, or another client) and the agent gets
searchandscrapeas first-class tools. Keyless means the server config has no secret to fill in. Browse the MCP server registry for how this compares to other search and browsing servers. - CLI scaffold.
npx -y firecrawl-cli@latest init --all --browserdrops config into your project. The--browserflag provisions the headless-browser path for JS-rendered pages;--allpulls the full toolset rather than a minimal subset. - Direct HTTP in a code harness. If your agent harness has code execution, call the endpoints directly and parse the Markdown — no MCP round-trips. This is the "code mode" pattern several teams now prefer for high-volume tool calls.
For document-heavy pipelines, pair it with Firecrawl's open-source pdf-inspector, which routes native-text PDFs around OCR.
Where a browser agent fits instead
Firecrawl is a fetch-and-clean layer. It is not the right tool for pages behind a login, flows that need clicks and form fills, or anything that has to act on your authenticated session. Those belong to a browser-automation agent — see Claude in Chrome and Claude's computer-use / browser tools. A common split: Firecrawl for open-web research, a browser agent for logged-in tasks.
Honest limitations
- Scraping at scale invites blocks and legal ambiguity. A keyless endpoint does not absolve you of
robots.txt, rate-of-fetch etiquette, or a site's ToS. Building a product on aggressive scraping of one source is fragile regardless of the tool. - Prompt-injection surface. Any page your agent scrapes can carry instructions aimed at the model. Clean Markdown does not strip adversarial text — see indirect prompt injection in AI agents. Treat scraped content as untrusted input.
- Free-tier lock-in risk is real but inverted. The danger isn't a bill — it's designing an agent around limits and latency that can change without notice, then scrambling when they do.
- Benchmark ≠ your workload. SimpleQA accuracy is encouraging; it is not evidence the tool wins on your query mix, your freshness needs, or your concurrency.
Bottom line
Firecrawl going keyless again is a meaningful lower of the barrier to giving an agent web search and scrape — no account, one command, working in under a minute. That is genuinely useful for prototyping, teaching, and one-off research.
The word to hold lightly is "free." Keyless removes signup, not rate limits, not the target web's terms, and not the absence of a reliability contract. Prototype on it freely; benchmark it against Exa and native model search on your queries; and before a customer-facing agent depends on it, move to an authenticated or self-hosted deployment you control.
Related on explainx.ai
- Firecrawl web scraping API for AI agents — the Agent endpoint and structured extraction, in depth
- Firecrawl pdf-inspector: PDF to Markdown without the OCR wait
- Perplexity Search API scores 80 in index debut, ahead of Parallel and Brave — the companion "which search API" half of the agent search/scrape pair
- OpenRouter web search benchmarks: how to pick a search tool for agents
- Perplexity's Search as Code: rethinking search for the agentic era
- What is MCP? Model Context Protocol architecture guide
- What is an agent harness? A complete guide
- Claude in Chrome: browser extension guide and safety
- Claude platform: computer use, browser, skills, Files API GA
- Indirect prompt injection in AI agents
- Publishers threaten to block crawlers as AI cuts referral traffic
Official
- Firecrawl announcement — Eric Ciarla (co-founder), August 28, 2026
- SimpleQA benchmark — OpenAI
Accurate as of August 29, 2026, based on Firecrawl's relaunch announcement. Rate limits, terms, and free-tier availability for keyless access are not fully published and can change — verify current limits and Firecrawl's terms of service before building on them, and check each target site's robots.txt and ToS. Follow @explainx_ai for agent-tooling updates.