Configuration · how to swap providers

The BSW Growth Agent is provider-agnostic by design. Same architecture, swap any layer for a free alternative or a different vendor. This guide shows you exactly where to swap and what to paste.

---

What you can swap

Layer	Default (in repo)	Free alternative	Premium alternative
Reasoning model (drafting)	Claude Sonnet 4.6	Groq · Llama 4 Scout	OpenAI GPT-5 (paid)
Routing model (classification)	Claude Haiku 4.5	Groq · Llama 3.1 8B Instant	Gemini 3 Flash
Web search	Claude web_search tool	Tavily free tier (1k/mo)	Perplexity Sonar
Web extraction	Firecrawl Cloud	Jina Reader (no key, free)	Browserbase (paid)
Orchestration runtime	n8n.cloud	n8n self-hosted Docker	LangGraph (code)
LLM hosting	Anthropic API	Ollama local (free)	Groq Cloud (free tier)

---

Recipe 1 · Zero-cost demo (Groq + Jina)

Best for: workshop attendees who don't want to enter a credit card. ~$0/month forever.

The drop-in lite workflow n8n/bsw-growth-agent-lite.json already implements this recipe end-to-end. Use that file unless you want to learn the swap mechanics yourself.

What changes

• LLM (drafting/discovery): swap claude-sonnet-4-6 and claude-haiku-4-5 for Groq's meta-llama/llama-4-scout-17b-16e-instruct (drafting/discovery) and llama-3.1-8b-instant (classification/digest)
• Web search: Groq has no built-in web search — use the public HN Algolia + Reddit JSON endpoints (both no-auth, no signup), or alternatively Tavily free tier (1,000 searches/month)
• Web extraction: swap Firecrawl for Jina Reader (no key, no signup)

Step 5 (Discovery) — Groq + Tavily

Replace the Anthropic HTTP Request node with two nodes:

Node 5a · Tavily search (HTTP Request)

• URL: https://api.tavily.com/search
• Method: POST
• Header: Authorization: Bearer YOUR_TAVILY_KEY
• Body:

{
  "query": "{{ $('Google Sheets').item.json.signal_keywords }}",
  "search_depth": "advanced",
  "max_results": 30,
  "include_domains": ["news.ycombinator.com", "reddit.com", "producthunt.com"]
}

Node 5b · Groq classify (HTTP Request)

• URL: https://api.groq.com/openai/v1/chat/completions
• Method: POST
• Header: Authorization: Bearer YOUR_GROQ_KEY
• Body:

{
  "model": "meta-llama/llama-4-scout-17b-16e-instruct",
  "messages": [
    {
      "role": "system",
      "content": "You are a customer-discovery research agent. Return ONLY a JSON array. No prose."
    },
    {
      "role": "user",
      "content": "From these search results: {{ JSON.stringify($('Tavily').item.json.results) }}\n\nExtract leads matching ICP: {{ $('Google Sheets').item.json.icp_description }}\n\nReturn JSON: [{ person, signal_type, source_url, evidence_quote, score 0-10 }]"
    }
  ]
}

Step 9 (Enrichment) — Jina Reader

Replace the Firecrawl HTTP Request node:

Old (Firecrawl):

URL: https://api.firecrawl.dev/v1/scrape
Header: Authorization: Bearer fc-...

New (Jina, no key):

• URL: https://r.jina.ai/{{ $json.company_url }}
• Method: GET
• (no auth header needed)
• Returns clean markdown directly

Step 11 (Drafting) — Groq Llama 4 Scout

Replace the Sonnet HTTP Request:

• URL: https://api.groq.com/openai/v1/chat/completions
• Body:

{
  "model": "meta-llama/llama-4-scout-17b-16e-instruct",
  "messages": [
    { "role": "system", "content": "[voice.md contents]" },
    { "role": "user",   "content": "Draft a customer-discovery email..." }
  ]
}

Quality note: Llama 4 Scout is fast (~1.5s/call on Groq, non-reasoning) and reliable for JSON extraction, but won't match Sonnet 4.6 for voice match. Expect ~10–15% more drafts to need editing before sending. For a free demo, fine. For production, upgrade to the paid workflow.

---

Recipe 2 · Local-first (Ollama)

Best for: anyone with a Mac M-series, Strix Halo, or RTX 4090+. Truly free after hardware.

Setup

1. Install Ollama: ollama.com/download
2. Pull models:

ollama pull qwen2.5:32b           # for drafting (replaces Sonnet)
ollama pull llama3.2:8b            # for classification (replaces Haiku)

3. Start Ollama: ollama serve

Workflow changes

Replace the Anthropic HTTP nodes with calls to your local Ollama:

• URL: http://localhost:11434/api/chat
• Method: POST
• Body:

{
  "model": "qwen2.5:32b",
  "messages": [...],
  "stream": false
}

If self-hosting n8n on the same machine: use http://host.docker.internal:11434 if n8n is in Docker.

Web search

Ollama has no built-in web search. Use Jina Reader for extraction and skip discovery (use a Reddit RSS feed as input), OR pair with Tavily free tier.

---

Recipe 3 · OpenRouter (one key, all models)

Best for: swappable model choice, low cost (~$10/mo at workshop scale).

Setup

1. Sign up at openrouter.ai
2. Get key from openrouter.ai/keys
3. Top up $10

Workflow changes

Same call structure, just change the URL and model name:

• URL: https://openrouter.ai/api/v1/chat/completions
• Header: Authorization: Bearer YOUR_OR_KEY
• Body:

{
  "model": "anthropic/claude-sonnet-4-6",
  "messages": [...]
}

Why it's nice: swap anthropic/claude-sonnet-4-6 for openai/gpt-5, meta-llama/llama-4-scout, google/gemini-2.5-pro etc. without changing anything else.

---

Recipe 4 · Gemini Free Tier (Google)

Best for: generous free limits (1,500 requests/day on Gemini 3 Flash).

Setup

1. Get key from aistudio.google.com/apikey
2. Free tier: 1,500 requests/day, 1M tokens/min

Workflow changes

Replace Anthropic HTTP nodes:

• URL: https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=YOUR_KEY
• Body uses Gemini's slightly different format · see Gemini API docs

Quality note: Gemini 3 Flash is fast and cheap. Voice match isn't quite Sonnet level but close.

---

How to share your forked repo without leaking keys

Three rules:

1. Use placeholders in the JSON · the workflow file has REPLACE_ME strings where credentials go. Real keys live in n8n's encrypted credential store, never in the file.

2. Never commit .env · only .env.example (which has REPLACE_ME values).

3. Run a key scan before pushing:

# Look for accidentally-committed secrets
grep -r "sk-ant-" . --include="*.json" --include="*.md" 2>/dev/null
grep -r "fc-[a-zA-Z0-9]" . --include="*.json" --include="*.md" 2>/dev/null
grep -r "sk-or-" . --include="*.json" --include="*.md" 2>/dev/null

# All three should return nothing.

If a key sneaks through, rotate it immediately — assume it's compromised the moment it hits a public repo.

---

Decision tree · which recipe should you pick?

Are you running this for personal use, or sharing publicly?

  Personal use ──→ Have ~$20/month to spend?
                      Yes ──→ Recipe 0 (default · Anthropic API · best quality)
                      No  ──→ Have a Mac M-series or RTX 4090+?
                                Yes ──→ Recipe 2 (Ollama local · free)
                                No  ──→ Recipe 1 (Groq + Jina · free, lower quality)

  Sharing publicly (workshop attendees, blog readers) ──→ Recipe 1 (Groq + Jina)
    (You don't want to require people to enter a credit card to try it)

  Want easy model-swapping in production? ──→ Recipe 3 (OpenRouter)

  Already deep in Google Cloud? ──→ Recipe 4 (Gemini)

---

What stays the same across all recipes

The architecture doesn't change:

• Trigger (cron · webhook · manual)
• Read config (Sheets + Drive)
• Discovery (whichever LLM)
• Dedup against state
• Enrich top-N (whichever extractor)
• Draft (whichever LLM)
• HITL gate (Gmail draft, never sends)
• Daily digest
• Audit log

Only the vendor of each step changes. Patterns don't change. Tools do.

---

Resources

• Default workflow: n8n/bsw-growth-agent.json — Anthropic + Firecrawl
• Lite workflow: n8n/bsw-growth-agent-lite.json — Groq + Jina (free tier path)
• Tutorial: TUTORIAL.md — full 17-step build
• Workshop deck: slides/index.html

---

MIT licensed · fork freely · ship yours · — Sophia Stein, AI Architect Questions? sophia@agenticarchitect.ai · agenticarchitect.ai/blog