Skip to main content
Firecrawl MCP exposes tools for finding, extracting, interacting with, and monitoring web content. Your MCP client receives the exact input schema for every available tool when it connects.

Tool availability

Start with Get Started and pick For Agents or For Humans. Some optional tools can be disabled by environment or team policy.

Choose a tool

firecrawl_gov_search also accepts the optional filters jurisdiction, court, courtType, cites, and domain (firecrawl-mcp 3.34.0 or later). See Government Index filters. jobs_search, firecrawl_save_skill, and firecrawl_research_lookup_protein are on the hosted server. The local firecrawl-mcp 3.40.0 release does not include them yet.
The former Extract MCP tool is deprecated and is not part of the current tool surface. Use Scrape for a known page and extract the fields from its content, or Agent when Firecrawl must discover the sources. See Choosing the Data Extractor for the full comparison.
Use the schema shown by your MCP client for the current arguments. The feature guides below explain the underlying Firecrawl behavior without duplicating those schemas here.

Important behavior

A local MCP server connected to a self-hosted Firecrawl API can read filePath directly. The hosted server cannot read files from your machine, so it uses a two-call handoff:
  1. Call firecrawl_parse with filePath to receive an upload command and uploadRef.
  2. Run the upload command on the machine that can read the file.
  3. Call firecrawl_parse again with the returned uploadRef.
The upload command uses a short-lived signed target and does not contain your Firecrawl API key. Use firecrawl_scrape for a public document URL.
firecrawl_crawl normally starts a crawl and polls it to a terminal state before returning. If that wait times out, resume the job with firecrawl_check_crawl_status and the crawl ID. Use the same status tool for a crawl created outside the current MCP call.firecrawl_agent is asynchronous: it returns a job identifier, and firecrawl_agent_status checks that job until it completes or fails.
Start with a url, or reuse the scrapeId from a previous Scrape call. When the workflow is finished, call firecrawl_interact_stop with the scrapeId to release the session.
The firecrawl_monitor_* family creates, lists, updates, runs, and inspects recurring monitors. firecrawl_monitor_delete permanently removes a monitor and should be called only when the user explicitly intends to delete it.
Alexandria tools need an authenticated session. Keyless sessions get no Alexandria tools.
  1. Discover. Authenticated firecrawl_search uses sources: ["web", "alexandria"] by default and returns matching tools in data.tools. Use sources: ["alexandria"] for tools only, or sources: ["web"] to omit semantic tool discovery.
  2. Explore. alexandria_library_search runs one read-only shell command (ls, cat, grep, jq, pipes) over the catalogue. Start with cat /library/README.md. Each tool has a contract at /library/<category>/<provider>/<capability>.md with its inputs, outputs, and the call that runs it, and recorded responses are under /examples. For a business workflow such as sales prep, prospect lists, or market research, check /library/skills first: cat /library/skills/search/<a-few-words> ranks your team’s and Firecrawl’s step-by-step skills, your team’s first.
  3. Run. Call alexandria_library_read with calls, an array of one to ten {provider, capability, options} entries, which may name different providers. Results come back in data.alexandria in the same order, and each has data or an error. A result too large to return inline is retained, and the returned nextTool reads it in slices.
The hosted server no longer lists firecrawl_find_tools, and its firecrawl_scrape lists url only. Both still answer the calls they used to, so earlier clients and prompts that use firecrawl_find_tools or firecrawl_scrape with alexandria keep working. New integrations should use the two library tools.Discovery and library lookups are free. Each run costs the price listed on its tool. If a provider needs accepted terms, the tool returns THIRD_PARTY_DATA_TERMS_REQUIRED with requiresAction.url. Give that URL to an org admin, who accepts the terms in the Firecrawl dashboard. The agent must not accept terms. After the admin confirms, send the same call again.
people_company_search and people_company_enrich run Firecrawl’s provider waterfalls over Alexandria people and company data. They need an authenticated session and are not listed to keyless sessions.
  • Search. Set entity to person or company, then narrow with filters such as titles, seniority, company_domains, locations, industries, or headcount. Providers run one at a time, and only those that apply every filter you give. Rows never carry emails or phone numbers. limit is 1 to 100, and max_credits defaults to 1000.
  • Enrich. Send up to 10 records per call, people and companies together. People get a profile and companies get firmographics. Add contact_fields (work_email, personal_email, mobile) to find contact details for people; mobile lookups run only when you ask for them. For each field, Firecrawl tries providers in order of listed price and past success rate and stops at the first one that finds it.
In people_company_enrich, a provider that finds a field is charged its listed price, and a miss is free. Without max_credits, an enrich call is uncapped; set it to cap what one call can spend. To call one specific provider, use alexandria_library_read.
Skills are step-by-step instructions an agent follows, naming Alexandria capabilities as @provider/capability. They live under /library/skills, which alexandria_library_search reads. firecrawl_save_skill needs an authenticated session and is free.
  • Create. Call it without id and give name, description, and body, or copy a skill with from: <id> for /library/skills/firecrawl/<id>, or team:<slug> for /library/skills/team/<slug>. Fields you also give override the copy. The new skill is enabled and listed at /library/skills/team/<slug>.
  • Change. Pass id (team:<slug>) and baseVersion, the version: in the skill’s SKILL.md, to change its name, description, or body, or whether it is enabled or listed. If the skill changed since you read it, the call returns a conflict; read it again and reapply.
Firecrawl’s library skills cannot be changed; copy one instead. Skills cannot be deleted with this tool.
jobs_search needs an authenticated session. Give at least one of query, company, url, or urls:
  • query searches postings by role, with location, remote, and posted_within_days. Add group_by: "employer" to see who is hiring, or boards_only to search employers’ own boards only.
  • company, a name or domain, lists the company’s open roles from its own job board. Add query for one role there.
  • url reads one posting and returns its description. urls checks up to 10 postings in one call.
Each row has the same fields, including title, company, location, remote, posted_at, url, source, status, and checked_at. status is open when Firecrawl read the posting from the employer’s board or live posting page in this call, and unverified when it comes from a job site. A posting checked by URL can also be closed, not_found, or error. Each provider read is charged its listed price, and max_credits (default 200) caps the call.
Set FIRECRAWL_NO_SEARCH_FEEDBACK=1 to prevent firecrawl_search_feedback from being registered. Set FIRECRAWL_NO_ENDPOINT_FEEDBACK=1 to prevent firecrawl_feedback from being registered.

Ultrasearch answers

On the hosted server, firecrawl_search can answer a query with a record from an Alexandria data provider as well as web results. When a search clearly names exactly one person, one company, one company’s open jobs, one research paper, or one to three US-listed stocks, Firecrawl looks it up and adds one provider record in a separate answer field. The web results in data stay the same. Ultrasearch adds nothing in these cases:
  • the query names more than one person, company or paper, or more than three stocks;
  • the query asks to choose between stocks, for example AAPL or MSFT, or compares one stock with something else;
  • the query excludes something, for example with not, except or a -term operator;
  • the query asks not to look anything up;
  • a stock query names a market or country outside the US, or a fund, an index or a cryptocurrency;
  • a product name with a price word asks what the product costs, not for its parent company’s stock.
In those cases, and when the provider finds no match, the search returns its web results as usual, sometimes with one short note in agent_hints.

When it runs

  • Ultrasearch runs only on hosted connections that sign in or use an API key. It doesn’t run on keyless sessions, for teams with zero data retention, or on a local server.
  • It runs when sources is unset or empty, or when it includes web or alexandria. A sources list with news but neither web nor alexandria gets stock answers only, at the same price. A list of only images turns it off.
  • Stock news works on any plan with credits.
  • Agents don’t need an extra parameter. The server instructions tell connected agents what answer can hold and how it is billed.

Billing

The provider record costs what the same lookup costs your team through Alexandria, on top of the search’s usual credits. When more than one provider can answer, the cheapest one is used. If that price is above the route’s ceiling per search (20 credits for a person or company, 10 for jobs), the lookup is skipped and nothing is charged for it. Person and company lookups are normally billed only when a record is found. When a person lookup tries more sources, each source costs your team’s price for it only if it finds the person, and a source is tried only while the lookup stays within the 20-credit ceiling. A jobs lookup is billed per posting it returns, so a company with no open postings costs nothing. Paper lookups are free. A stock answer with headlines costs 2 credits on top of the search. A stock lookup that finds no headlines costs nothing extra, and the delayed price is always free. answer.credits shows what the search and the lookup cost.

The answer field

answer sits next to data in the search result. An illustrative example with fake data:
For stock news, prices holds a delayed price for each ticker and data lists the headlines:
The stock price comes from Fiscal.ai and is never real time. During the trading day it is the last trade, delayed by 15 minutes ("delay": "15-minute delayed"). Before the day’s first trade, and on days the market is closed, it is the last close ("delay": "end-of-day close"). as_of gives the time of that trade or close in New York time. If no price can be fetched, the answer carries the news alone; if no story is found but a price is, it carries the price with an empty data list.

Stock headlines

The number of headlines depends on what the query asks, per ticker: Headlines about the asked company come first, newest first, with the ones that fit the question, such as an earnings report or an analyst action, ahead. When there aren’t enough, the newest other stories Benzinga tags with the ticker fill the rest and carry "focus": "tagged". A price ask that has its price shows only headlines about the company, so it can answer with the price alone. The answer is held to a fixed size, so it may carry fewer headlines.

Turn it off

To search without person, company, jobs or paper lookups, pass a sources list without web or alexandria, such as ["news"]. A news-only search can still get a stock answer. A list of only images turns Ultrasearch off completely. There is no separate switch.

Feature guides

Scrape

Extract content or structured fields from one URL.

Search

Find relevant web, news, image, and developer sources.

Research Index

Search papers, read passages, and follow citations.

Developer Index

Answer coding questions from issues, PRs, READMEs, and docs.

Government Index

Search statutes, regulations, codes, and court opinions.

Crawl

Traverse and extract a site or section.

Parse

Convert files into LLM-ready output.

Interact

Operate dynamic pages in a live browser session.

Agent

Run autonomous multi-source research.

Monitoring

Track page changes and receive notifications.

Alexandria

Find and run data tools from third-party providers.

Troubleshooting

  • A tool is missing: confirm the connection mode on Get Started, reconnect or restart the client, and check whether team policy disables optional tools.
  • The client returns 401: check the configured server URL first.
    • If the configured URL is /v2/mcp-oauth, sign in again through the client.
    • If it is /v2/mcp, either replace the API key on that server or update the existing server URL to /v2/mcp-oauth and complete sign-in.
    • Start a new client session after either change.
  • The client is rate-limited: review the current rate limits, wait for the retry interval, or move from keyless to authenticated access.