Claude Agent SDK Cold Email: Build Claude-Native Outreach Agents with FoxReach

Why Claude Agent SDK for cold email

Claude's advantage for cold email is tone and context. The 200k-token window lets you feed an entire LinkedIn profile plus the prospect's last 20 blog posts plus three earnings calls into one message without summarization. Sonnet 4 produces the best default-tone copywriter in the current model lineup. Extended thinking gives measurably better openers on tone-sensitive outreach.

The Claude Agent SDK (anthropic Python SDK with agent primitives) is the lowest-abstraction way to use these features. No framework in between. You get new Anthropic releases (extended thinking, MCP support, prompt caching) on day one. If Claude is your primary model, skip LangChain and go direct.

Prerequisites

• Python 3.10+

  Anthropic SDK 0.40+ supports MCP and extended thinking.

• An Anthropic API key

  Generate at console.anthropic.com. Start with Sonnet 4 tier for development.

• A FoxReach API key

  Free plan works. Settings → API Keys.

Quick setup

Install the SDK and register the FoxReach MCP server. The SDK handles MCP tool discovery, tool use formatting, and streaming in the same API.

pip install anthropic foxreach

import os
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

message = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=2048,
 mcp_servers=[
 {
 "type": "url",
 "url": "https://api.foxreach.io/mcp",
 "name": "foxreach",
 "authorization_token": os.environ["FOXREACH_API_KEY"],
 },
 ],
 messages=[
 {
 "role": "user",
 "content": "Create a draft campaign called 'SaaS CTOs April' in FoxReach and add a 3-step sequence - intro, value prop, soft ask. 3 days between each.",
 },
 ],
)

for block in message.content:
 if block.type == "text":
 print(block.text)
 elif block.type == "mcp_tool_use":
 print(f"Tool: {block.name} Args: {block.input}")

Run it and Claude chains tool calls - campaigns.create, then three sequences.create - and reports the outcome. No agent framework in between, no tool wrappers to write. The MCP server handles tool discovery; Claude handles tool-use protocol negotiation.

Long context for research

The cheap move is to summarize research before handing it to Claude. Don't. Claude's 200k window is the advantage; using it fully produces noticeably sharper openers. Feed the raw LinkedIn profile, the recent blog posts, the earnings transcript excerpt, and the Crunchbase funding note as separate document blocks. Claude picks the right detail for the opener.

message = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=1024,
 messages=[
 {
 "role": "user",
 "content": [
 {"type": "text", "text": "Write a 60-word cold email to this person. Open with one specific, citable fact from the research below."},
 {"type": "text", "text": f"LinkedIn profile:\n{profile_text}"},
 {"type": "text", "text": f"Recent blog posts:\n{posts_text}"},
 {"type": "text", "text": f"Funding history:\n{crunchbase_text}"},
 {"type": "text", "text": f"Template to personalize:\n{template_text}"},
 ],
 },
 ],
)

Streaming with tool use

For user-facing agents (a Slack bot, a custom UI), stream the response. The Anthropic SDK streams text blocks, tool-use events, and thinking blocks separately - you can show typing while tool calls resolve in parallel.

with client.messages.stream(
 model="claude-sonnet-4-6",
 max_tokens=2048,
 mcp_servers=[{"type": "url", "url": "https://api.foxreach.io/mcp", "name": "foxreach", "authorization_token": api_key}],
 messages=[{"role": "user", "content": "Show me my active campaigns."}],
) as stream:
 for text in stream.text_stream:
 print(text, end="", flush=True)
 final = stream.get_final_message()
 for block in final.content:
 if block.type == "mcp_tool_use":
 print(f"\n[Called {block.name}]")

Prompt caching

For agents that run many campaigns against the same system prompt and tool definitions, enable prompt caching. Anthropic caches the cached blocks for 5 minutes by default; subsequent calls re-use them at ~10% of the full token cost. For a 5k-token system + tool definitions block called 100 times per hour, this is a real saving.

message = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=2048,
 system=[
 {
 "type": "text",
 "text": SYSTEM_PROMPT,
 "cache_control": {"type": "ephemeral"},
 },
 ],
 mcp_servers=[{"type": "url", "url": "https://api.foxreach.io/mcp", "name": "foxreach", "authorization_token": api_key}],
 messages=[{"role": "user", "content": user_request}],
)

Cache hits are reported in message.usage.cache_read_input_tokens. Log this in production to confirm caching is actually landing.

Common pitfalls

Frequently asked questions