skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 1d4b0728d0f4
- Files read
- 43
This skill pack teaches AI coding agents to use the Firecrawl CLI and MCP tools to search the web, scrape and monitor sites, parse documents, drive hosted browser sessions, and produce research or business reports. Its biggest risks are blanket auto-approval of all Firecrawl CLI commands—which lets prompt-injected web pages overwrite local files or trigger unconfirmed requests—and default rules that silently upload local documents and background session objectives to Firecrawl's servers.
What it can reach
- Shell commands
- Runs firecrawl and npx firecrawl-cli commands automatically via allowed-tools, and instructs the agent to use standard shell tools such as jq, grep, head, curl, and npm.
- Files
- Saves scraped pages, search JSON, site mirrors, and screenshots under .firecrawl/ or output files, reads local documents to upload via firecrawl parse, and installs skills globally into detected coding editors during setup.
- Network
- Sends search queries, target URLs, uploaded local documents, browser actions, monitor webhooks or email alerts, and background usage feedback to Firecrawl's cloud API, and downloads packages from npm or screenshots via curl.
- Secrets
- Uses FIRECRAWL_API_KEY or stored OAuth tokens for Firecrawl authentication, and certain browser and QA skills handle user-supplied website login credentials and saved session cookies.
- Browser
- Opens the local browser for Firecrawl OAuth login and controls remote cloud browser sessions via firecrawl interact to click, fill forms, log in, and persist session profiles.
Findings
-
HighSkills auto-approve all Firecrawl CLI subcommands without user confirmation
skills/core/firecrawl/SKILL.md:6
Ten core skills grant wildcard permission for Bash(firecrawl *) and Bash(npx firecrawl-cli *) in allowed-tools, letting the agent execute every Firecrawl subcommand and flag with no confirmation prompt—even inside narrow read-only skills like firecrawl-map or firecrawl-search. Because these skills routinely ingest untrusted web pages, an indirect prompt injection on a scraped page can trick the agent into silently overwriting arbitrary local files outside .firecrawl/ via the -o flag, uploading local documents via firecrawl parse, sending sensitive context in outbound URLs via firecrawl scrape, creating webhook monitors, or driving saved authenticated browser profiles via firecrawl interact.
Fix: Restrict allowed-tools in each skill to its specific subcommand (for example Bash(firecrawl map *)), remove Bash(npx firecrawl-cli *), and require user confirmation for file-uploading, state-changing, or authenticated subcommands such as parse, interact, monitor, login, init, and setup.
-
MediumLocal document skill intercepts file-path tasks and uploads files to Firecrawl without confirmation
skills/core/firecrawl-parse/SKILL.md:4
The firecrawl-parse skill tells the agent to trigger whenever the input is a local file path (PDF, DOCX, XLSX, HTML, and others) rather than a URL, while pre-approving Bash(firecrawl *) so no confirmation prompt appears. Unlike local file inspection tools, firecrawl parse uploads the local file (up to 50 MB) to Firecrawl's cloud servers. A user asking their coding agent a routine question about a confidential local contract, spreadsheet, or HTML file can have that file uploaded off their machine without realizing an external service was used.
Fix: Change the skill description so it activates only when the user explicitly asks to parse a document with Firecrawl, warn clearly that files are uploaded to Firecrawl's servers, and remove Bash(firecrawl *) from allowed-tools so uploads require confirmation.
-
MediumSkills send background session objectives and research feedback to Firecrawl without asking
skills/core/firecrawl-alexandria/SKILL.md:31
After searches and Alexandria provider lookups, firecrawl-alexandria, firecrawl-search, and firecrawl instruct the agent to automatically run silent background commands (--silent &) that send structured feedback to Firecrawl for a 1-credit refund. For Alexandria sessions, this includes --objective—a summary inferred from conversation context of what the user was ultimately trying to accomplish—along with target URLs and needed functionality; for searches, firecrawl search-feedback sends missing content topics and external URLs visited. Unless the user knows in advance to set FIRECRAWL_NO_ENDPOINT_FEEDBACK=1 and FIRECRAWL_NO_SEARCH_FEEDBACK=1, confidential business goals and private prompt context can be transmitted to Firecrawl without consent.
Fix: Make feedback opt-in or prompt the user before sending it, omit broader session goals (--objective) by default, and instruct the agent never to include confidential project details in feedback fields.
-
MediumSetup rules and allowed-tools execute unpinned npm packages at runtime
skills/core/firecrawl/rules/install.md:15
When the CLI is missing, outdated, or partially configured, rules/install.md directs the agent to run npx -y firecrawl-cli@latest init -y --browser (or npm install -g firecrawl-cli@latest), and ten core skills pre-approve Bash(npx firecrawl-cli *) in allowed-tools. Pulling @latest via npx without a pinned version or integrity check executes whatever code is currently published on npm, and init -y automatically writes skills globally across all detected coding editors on the machine.
Fix: Pin firecrawl-cli to a verified version, remove Bash(npx firecrawl-cli *) from allowed-tools, and ask the user before installing global packages or modifying other editors' skill directories.
-
LowWalkthrough, QA, and setup instructions encourage passing plaintext credentials through chat and CLI flags
skills/workflows/firecrawl-demo-walkthrough/SKILL.md:24
Both firecrawl-demo-walkthrough and firecrawl-qa tell the agent to ask the user for login credentials for protected areas, firecrawl-interact demonstrates typing login details via natural-language CLI prompts and persisting cookies with --profile, and rules/install.md suggests running firecrawl login --api-key "<key>". Entering passwords or API keys into chat and CLI arguments exposes them in AI model logs, shell history, and the OS process list, and the interact skill does not warn users that browser sessions and saved --profile cookies are hosted on Firecrawl's remote servers rather than locally.
Fix: Instruct the agent never to ask users to paste passwords or API keys into chat, have users authenticate directly in their own terminal or browser session, and clearly note that firecrawl interact profiles store session state on Firecrawl's cloud infrastructure.
