Skip to content

README.md

Latest commit

 

History

History
194 lines (136 loc) · 9.01 KB

README.md

File metadata and controls

194 lines (136 loc) · 9.01 KB

Experimental: AI Workflows

Launch pre-built AI workflows powered by Firecrawl + your coding agent. One command gathers your inputs, then drops you into an interactive agent session with the right tools and instructions.

Think ollama run but for web research agents. All workflows spawn parallel subagents to divide the work and finish faster.

Supported Backends

Backend Command Status
Claude Code firecrawl claude <workflow> Available
Codex (OpenAI) firecrawl codex <workflow> Coming soon
OpenCode firecrawl opencode <workflow> Coming soon

The same workflows work across all backends - only the spawn command changes. System prompts, interactive flows, and output formats are shared.

Prerequisites

  • Firecrawl CLI installed and authenticated (firecrawl login)
  • At least one supported backend installed:
    • Claude Code: npm i -g @anthropic-ai/claude-code
    • Codex: npm i -g @openai/codex
    • OpenCode: install instructions

Available Workflows

competitor-analysis - Competitive landscape analysis

Spawns parallel agents -- one per company -- to scrape and profile the target and each competitor simultaneously. The team lead synthesizes results into a full report: company overviews, feature comparison matrix, positioning analysis, strengths/weaknesses, and actionable recommendations. Citations included by default.

# Interactive (asks questions step by step)
firecrawl claude competitor-analysis

# One-liner
firecrawl claude competitor-analysis https://firecrawl.dev -y

# With specific competitors and JSON output
firecrawl claude competitor-analysis https://firecrawl.dev \
  --competitors "apify,scrapy,crawlee" \
  -o json -y
Option Description
--competitors <list> Comma-separated competitor names or URLs
--context <text> Additional context for the analysis
-o, --output <format> terminal (default), markdown, json
-y, --yes Auto-approve all tool permissions

deep-research - Multi-source topic research

Breaks the topic into research angles, then spawns parallel agents -- one per angle (overview, technical, market, contrarian). Each agent searches and scrapes from their perspective. Results are cross-referenced and synthesized into a structured report. Three depth levels control how many sources are consulted.

# Interactive
firecrawl claude deep-research

# One-liner with depth
firecrawl claude deep-research "AI code generation landscape" --depth exhaustive -y

# Quick overview saved to file
firecrawl claude deep-research "WebSocket alternatives" --depth quick -o markdown -y
Option Description
--depth <level> quick (5-10 sources), thorough (15-25), exhaustive (25+)
--context <text> Specific angles or questions to focus on
-o, --output <format> terminal (default), markdown, json
-y, --yes Auto-approve all tool permissions

lead-research - Pre-meeting intelligence brief

Spawns parallel agents to research a company, recent news/activity, and optionally a specific person -- all at once. Results are synthesized into a brief with talking points and pain points.

# Interactive
firecrawl claude lead-research

# One-liner
firecrawl claude lead-research "Stripe" --person "Patrick Collison" --context "partnership call" -y
Option Description
--person <name> Specific person to research
--context <text> Meeting context (e.g., "sales call", "partnership eval")
-o, --output <format> terminal (default), markdown
-y, --yes Auto-approve all tool permissions

seo-audit - Full SEO audit

Maps the site, then spawns parallel agents for site structure, on-page SEO, and keyword/competitor analysis. Produces a prioritized audit with specific (not generic) recommendations.

# Interactive
firecrawl claude seo-audit

# One-liner
firecrawl claude seo-audit https://example.com --keywords "scraping,api,web data" -y
Option Description
--keywords <list> Comma-separated target keywords
-o, --output <format> terminal (default), markdown
-y, --yes Auto-approve all tool permissions

qa - Parallel QA testing with cloud browser

Acts as a QA team lead: maps the site, then spawns 3-4 parallel subagents that use Firecrawl's cloud browser to click around, fill forms, test interactions, and find bugs simultaneously. Results are merged into a unified report.

# Interactive
firecrawl claude qa

# One-liner
firecrawl claude qa https://myapp.com -y

# Focus on specific area
firecrawl claude qa https://myapp.com --focus forms -y
Option Description
--focus <area> full (default), forms, navigation, responsive, performance
--context <text> Specific areas or known issues to check
-o, --output <format> terminal (default), markdown
-y, --yes Auto-approve all tool permissions

Focus areas and their agent teams:

Focus Agents
full Navigation & Links, Forms & Interactions, Content & Visual, Error States
forms Form Discovery, Happy Path, Edge Cases, Validation
navigation Sitemap, Nav Testing, Link Checker, Routing
responsive Desktop, Tablet, Mobile, Interaction
performance Page Load, Asset Audit, Content Efficiency, Comparison

How It Works

Each workflow:

  1. Gathers inputs - Interactive prompts (or CLI flags for one-liners)
  2. Asks about permissions - Auto-approve all tools, or approve each manually
  3. Launches the agent - Spawns an interactive session with workflow instructions as the system prompt
  4. Agent self-discovers tools - First thing it does is run firecrawl --help to learn the full CLI
  5. Parallel subagents - The lead agent spawns specialized subagents that work simultaneously
  6. You stay in control - It's an interactive session, so you can follow up, redirect, or ask for more detail

The -y Flag

By default, each workflow asks how the agent should handle tool permissions:

? How should the agent handle tool permissions?
> Auto-approve all (recommended)
  Ask me each time

Pass -y to skip this prompt and auto-approve everything. This adds --dangerously-skip-permissions to the Claude spawn.

Adding New Workflows

Each workflow is a function pair in claude.ts:

  1. gatherXxxInputs() - Interactive prompts using @inquirer/prompts
  2. buildXxxSystemPrompt() - System prompt with instructions and output format

The shared launchAgent() function handles spawning. Add a new command in registerWorkflows() and you're done.

Multi-Backend Architecture

The system prompts and interactive flows are backend-agnostic - they just tell the agent to use firecrawl CLI commands. The only backend-specific part is the spawn function:

Backend Spawn System prompt Skip permissions
Claude Code claude --append-system-prompt --dangerously-skip-permissions
Codex codex --config / profile --full-auto or --yolo
OpenCode opencode run --prompt OPENCODE_PERMISSION env var

To add a new backend, create a launcher function and register the command group in index.ts.