yente-client

Python SDK for the OpenSanctions API and on-premise yente instances.

Full documentation, including the tutorial and API reference, lives at yenteclient.followthemoney.tech.

Install

pip install yente-client            # SDK only
pip install 'yente-client[cli]'     # SDK + `yente-cli` command-line tool
pip install 'yente-client[mcp]'     # SDK + `yente-mcp` server for LLM agents

Python 3.11+; runtime deps are pydantic and httpx. The [cli] extra adds typer and rich; the [mcp] extra adds fastmcp. The MCP server exposes the matching surface to LLM agents over the Model Context Protocol.

Quickstart

from yente_client import Client, Person

with Client(api_key="...", app_name="MyScreeningApp") as c:
    hits = c.match(
        Person(firstName="Aleksandr", lastName="Zacharov", birthDate="1965"),
        datasets=["sanctions"],
        threshold=0.7,
    )
    if hits.top is not None:
        print(hits.top.caption, hits.top.score)
    for match in hits.matches:
        print(match.id, match.properties.get("topics", []))

The API key can be generated at opensanctions.org/account. It's read in this example from the OPENSANCTIONS_API_KEY env var if you skip passing api_key=. To target a yente instance, pass base_url= (no key needed).

Other endpoints

# Free-text search
res = c.search("acme", datasets=["default"], schema="Company")

# Fetch one entity by ID; nested=True (default) inlines adjacent entities
entity = c.fetch("NK-aU5ybkbRFJucf8YMwsJvDw")
for sanction in entity.properties.get("sanctions", []):
    print(sanction.properties["authority"])

# Operational endpoints
c.datasets()      # available datasets and freshness
c.programs()      # sanctions-program catalog (resolves programId codes)
c.algorithms()    # enabled matching algorithms
c.healthz()       # liveness

Entity construction

The package ships generated classes for every FtM schema (Person, Company, Vessel, Organization, …). All take typed list[str] properties; a single string is coerced to a one-element list. Unknown properties raise pydantic.ValidationError at construction.

from yente_client import Person, Company

p = Person(firstName="Aleksandr", lastName="Zacharov", country="ru")
c = Company(name="Acme LLC", jurisdiction="us")

Person(birth_date="1965")    # ValidationError — snake_case isn't aliased
Person(notARealProp="X")     # ValidationError — extra="forbid"

Configuration

Client accepts:

CLI

pip install yente-client[cli] ships a yente-cli binary that mirrors the SDK:

export OPENSANCTIONS_API_KEY=sk_...        # or pass --api-key

# Screen a known entity (KYC / sanctions checks):
yente-cli match -s Person -p firstName=Aleksandr -p lastName=Zacharov -d sanctions

# Free-text discovery by name:
yente-cli search "acme" -d default -s Company

# Fetch one entity (id from match/search):
yente-cli fetch NK-aU5ybkbRFJucf8YMwsJvDw

# Discover the data model (offline, no API key):
yente-cli ref schemas                   # all schemas with matchable flags
yente-cli ref schema Person -f json     # full property list, types, deprecation
yente-cli ref topics                    # the Topic enum
yente-cli ref countries                 # country codes the server speaks

# Discover server state:
yente-cli status                        # client + server + auth + loaded datasets
yente-cli datasets                      # full per-source dataset list
yente-cli programs                      # sanctions programs behind programId codes
yente-cli algorithms                    # enabled algorithms, default + best

Output formats: -f table (default on TTY), -f json (pretty, default when piped), -f jsonl (one item per line, ideal for jq and LLM pipelines).

search vs match: use match for any matching task, even with partial input (a name, name + country, …) — it returns scored, ranked candidates. search is for user-facing search UIs (a search box or autocomplete a human types into), not a fallback for match on sparse input.

Exit codes:

• 0 ≥1 result
• 1 zero results (lets shell scripts gate on &&)
• 2 usage error (bad flag, unknown schema/property)
• 3 API error (4xx, 5xx)
• 4 network/transport error

Designed for LLM agents: every command's --help carries worked examples and documented JSON output shapes; unknown schema/property names get fuzzy suggestions ("Did you mean birthDate?"). Run yente-cli --help first.

Errors

Every non-2xx response raises a subclass of YenteError:

• BadRequestError (400)
• AuthenticationError (401, 403)
• NotFoundError (404)
• RateLimitError (429, with .retry_after when set)
• ServerError (5xx)
• APIError (other; carries .status_code and .detail)
• TransportError (network failure before the request reached the server)

Retries are not built in — failed requests raise; callers handle backoff.