iwantfyi-langchain

LangChain tools for iwant.fyi — the reference implementation of the iwant.fyi demand-side protocol v1.0.

iwant.fyi demand-side protocol is an open standard for how AI agents express structured purchase intent on behalf of users, receive matched supply across multiple sources, and report outcomes back. This package wraps iwant.fyi as a set of LangChain StructuredTool objects that any LangChain agent can use.

Install

pip install iwantfyi-langchain

Quick start

from iwantfyi_langchain import get_iwant_tools
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

tools = get_iwant_tools(api_key="iwant_ak_...")  # get a key at https://iwant.fyi

agent = create_react_agent(
    ChatOpenAI(model="gpt-4o-mini"),
    tools,
)

result = agent.invoke({
    "messages": [{"role": "user", "content": "Find me a 1/4-inch drive torque wrench under $150"}]
})
print(result["messages"][-1].content)

That's it. The agent now has seven tools wired in: demand_create_want, demand_search, demand_get_want, demand_record_outcome, demand_list_verticals, demand_list_constraints, demand_health.

Get an API key

1. Visit iwant.fyi and sign in.
2. Register an agent (UI or POST /api/agents).
3. Save the returned key (format: iwant_ak_...).

What you get

All tools use Pydantic v2 args schemas, so LangChain will validate user inputs before invoking iwant.fyi.

Direct client (no LangChain)

You can use the underlying IwantClient directly if you don't want LangChain in the loop:

from iwantfyi_langchain import IwantClient

with IwantClient(api_key="iwant_ak_...") as client:
    health = client.health()
    print(health["status"])

    response = client.create_want(
        title='Torque wrench, 1/4" drive, 25-100 ft-lb',
        price_cents=15000,
        vertical="tools",
        mode="any",
    )
    print(f"Want {response['want']['id']} -> {response['matches']['match_count']} matches")

Transports

By default the client uses MCP over HTTP (JSON-RPC). Switch to the REST fallback:

client = IwantClient(api_key="...", transport="http")

Both transports cover all demand.* tools. Legacy iwant.fyi tools (browse_wants, search_listings, etc.) only work over MCP transport via client.call_tool(...).

Outcome events

Outcome events feed match-quality learning and (eventually) revenue-share attribution:

client.record_outcome(want_id=w_id, match_id=m_id, event="clicked")
client.record_outcome(want_id=w_id, match_id=m_id, event="purchased", value_cents=12500)

Events are idempotent on the server side. Replays are no-ops.

Errors

Typed exceptions from iwantfyi_langchain.errors:

from iwantfyi_langchain import IwantError, UnauthorizedError, ValidationError, RateLimitedError

try:
    client.create_want(title="x", price_cents=1)
except UnauthorizedError:
    ...  # refresh key
except ValidationError as e:
    ...  # e.message, e.data
except RateLimitedError:
    ...  # back off
except IwantError as e:
    ...  # generic; e.code, e.message

Specification

Full iwant.fyi demand-side protocol v1.0 specification: iwant.fyi/protocol/v1

The protocol is published under Apache 2.0. iwant.fyi is the reference implementation; anyone may build their own.

License

Apache 2.0