langchain-layrcake

LangChain tools for Layrcake, go-to-market infrastructure for AI agents. Your agent can find leads, find and verify work emails, check and research companies, and draft email campaigns that a person approves before anything sends.

The tools wrap the official layrcake Python SDK. They work with any LangChain or LangGraph agent that calls tools.

Install

pip install -U langchain-layrcake

Get an API key at layrcake.dev/connect and set LAYRCAKE_API_KEY. New workspaces start with 300 free credits.

Quickstart

This example also needs LangChain and a model provider: pip install -U "langchain[anthropic]".

from langchain.agents import create_agent
from langchain_layrcake import LayrcakeToolkit

toolkit = LayrcakeToolkit(budget_credits=50)  # reads LAYRCAKE_API_KEY
agent = create_agent(model="anthropic:claude-sonnet-5-5", tools=toolkit.get_tools())

result = agent.invoke(
    {
        "messages": [
            {
                "role": "user",
                "content": (
                    "Find 5 heads of sales at UK software companies, verify their work emails, "
                    "draft a two-step campaign and ask for approval."
                ),
            }
        ]
    }
)
print(result["messages"][-1].content)

The agent stops at an approval link. Nothing is sent until a person approves it in the Layrcake console.

Try it free first: LayrcakeToolkit(dry_run=True) makes every call return its cost instead of data and charges nothing.

A test run

On 20 October 2026 we ran this quickstart with the model openrouter:anthropic/claude-sonnet-5.5 (through langchain-openrouter), budget_credits=25 and this prompt: "Find 3 heads of sales at software companies in London with 20 to 200 staff, verify their work emails, and draft a short campaign about cutting time spent on lead research."

The agent called get_me, find_leads, enrich three times and verify_email twice. One person had no work email, so it called find_leads, enrich and verify_email once more for a replacement, then create_campaign and launch_campaign. It found 6 people, kept 3 verified emails (1 catch-all) and spent 24.75 of its 25 credits. It stopped at the approval link, and nothing was sent. The run took 67 seconds. The draft emails were signed "[Your name]", so edit the sign-off before you approve.

Tools

1 credit is about $0.01. New workspaces start with 300 free credits. Current prices are at docs.layrcake.dev/pricing.html.

Use the whole toolkit, or single tools. Tools built from the same LayrcakeSession share one run:

from langchain_layrcake import LayrcakeFindLeads, LayrcakeSession, LayrcakeVerifyEmail

session = LayrcakeSession(budget_credits=20, dry_run=True)
tools = [LayrcakeFindLeads(session=session), LayrcakeVerifyEmail(session=session)]

print(tools[1].invoke({"email": "ada@example.com"}))
# {"dry_run": true, "would_cost_up_to": "1.25", "would_exceed": [], "warnings": []}

How approval works

Nothing reaches a real person until a person on your side says yes.

1. create_campaign saves a draft. A draft never sends.
2. launch_campaign asks for approval. It returns status: "needs_confirmation", a preview of what would send and an approval_url.
3. A signed-in person opens the link and approves in the Layrcake console. The Layrcake API enforces this step.
4. Only then does the campaign start sending.

The launch_campaign tool has no confirmation_token argument, and the confirmation token is removed from what the model reads, so an agent cannot approve its own send. After launch_campaign asks for approval, the other paid tools in the same run refuse to run, and the agent should stop and report the link.

Run rules

The tools of one run share a LayrcakeSession from layrcake.agent_tools, the one implementation of these rules in the Layrcake Python SDK. It applies the rules from layrcake.dev/agents.md in code:

• Budget. Every paid call sends what is left of the run budget as budget_credits, so the API refuses a call that could cost more. Once the budget is spent, paid tools refuse to run. The default budget is 50 credits.
• Dry run. With dry_run=True, every call returns its plan instead: the most it could cost and any limit it would trip. Nothing is looked up or charged.
• Idempotency. Every call sends an idempotency key derived from the run and its arguments. A repeated call returns the first answer instead of charging again. Pass the same run_id to resume a run safely.
• Verified only. create_campaign refuses an address that did not pass verify_email in this run (valid or catch_all_valid). Turn this off with require_verified=False if you verify elsewhere.
• Errors as text. API errors and refusals come back to the model as text that starts with Refused: and says why, such as a spent budget or missing credits, so it can fix the call or stop.

Every call goes through the official layrcake Python SDK, which handles authentication, retries and typed errors.

Configuration

Use one toolkit per agent run. Before a live launch_campaign, connect sending mailboxes in the Layrcake console; without them the call returns an error that says what is missing.

Development

uv sync
uv run ruff check . && uv run ruff format --check .
uv run mypy
uv run pytest tests/unit_tests --disable-socket --allow-unix-socket
LAYRCAKE_API_KEY=... uv run pytest -m live tests/integration_tests   # dry runs only, charges nothing

Unit tests run against a mocked API and include LangChain's standard tool tests from langchain-tests. The run rules are tested in the SDK.

Links

• Layrcake docs: docs.layrcake.dev
• For agents: layrcake.dev/agents.md
• MCP server: https://mcp.layrcake.dev/mcp
• Python SDK: github.com/Layrcake-dev/layrcake-python

License

MIT