Nimble proof of concept for Tangos investigations

Your code calls Nimble at three steps of an investigation

• Each step runs during the investigation, when your pipeline needs the data.
• Nothing is harvested in bulk, and nothing is stored after the run.
• Every value comes back with the URL it was read on.

Two agents are ready in your Nimble account

Log in to your Nimble account to open them. Each one already has two runs on your sample cases.

We score each use case on your own cases

One case is one call from your code. We compare what that call returns with what your team found.

We need four things from your team

• The Danske Bank and Boeing findings that mattered in your own investigations.
• The identifiers you hold on Gilbert Verdian and Brad Jackson, such as date of birth, employer or ID number.
• About 10 more cases per use case, each with the result you expected.
• The search-page links that give you the most trouble today.

The trial ends with scores on your cases in the week of 26 October

Your LLM drives Nimble Extract through any search form

• Some sanctions and registry links open a search page instead of the record.
• A script per site breaks when the site changes, and you cannot write one for a site you have not seen.
• Nimble Extract opens the page in a real browser. It can type into a field, click a button and wait for the result.
• Your LLM reads each page and chooses the next action, so no site needs its own code.

Your LLM runs the same five steps on every site

Both of your sample links reach the record in three Extract calls

We ran these calls on 12 October 2026 on the two links your team sent. Each block shows what Extract returned and what the LLM decides next.

<input name="ctl00$MainContent$txtLastName" title="Enter name as search criteria.">
<input name="ctl00$MainContent$txtAddress" title="Enter part of an address as search criteria.">
<input name="ctl00$MainContent$txtID" title="Enter ID number or digital currency address as search criteria.">
<select name="ctl00$MainContent$ddlType" title="Select an SDN Type">
<select name="ctl00$MainContent$ddlCountry" title="Select a country as search criteria.">
<input type="submit" name="ctl00$MainContent$btnSearch" value="Search">

[
  {"fill": {"selector": "#ctl00_MainContent_txtLastName", "value": "Maduro"}},
  {"click": {"selector": "#ctl00_MainContent_btnSearch", "timeout": 8000}},
  {"wait_for_element": {"selector": "#gvSearchResults", "visible": true, "timeout": 15000}}
]

Lookup Results: 4 Found
FELIX B. MADURO S.A.                 Details.aspx?id=19147
FLORES DE MADURO, Cilia Adela        Details.aspx?id=25079
MADURO GUERRA, Nicolas Ernesto       Details.aspx?id=26946
MADURO MOROS, Nicolas                Details.aspx?id=22790

Type: Individual            List: SDN
Last Name: MADURO MOROS    Program: VENEZUELA; IRAN-CON-ARMS-EO
First Name: Nicolas
Title: President of the Bolivarian Republic of Venezuela
Date of Birth: 23 Nov 1962  Place of Birth: Caracas, Venezuela
Citizenship: Venezuela
Cedula No.: 5892464 (Venezuela)

Search UN Security Council Consolidated Lists  -> https://search.sanctions.un.org
HTML  -> https://scsanctions.un.org/consolidated
XML   -> https://scsanctions.un.org/resources/xml/en/consolidated.xml

[
  {"wait_for_element": {"selector": "input.MuiInputBase-input", "visible": true, "timeout": 20000}},
  {"click": {"selector": "input.MuiInputBase-input", "timeout": 5000}},
  {"fill": {"selector": "input.MuiInputBase-input", "value": "Dawood Ibrahim", "mode": "type"}},
  {"click": {"selector": "button.search-btn", "timeout": 8000}}
]

Total: 51
Score  Full name   Type        Reference  Regime
12.51  DAWOOD      Individual  QDi.135    Al-Qaida
 8.82  IBRAHIM     Individual  QDi.299    Al-Qaida
 6.17  IBRAHIM     Individual  IQi.084    Iraq
 5.86  DAWOOD      Individual  IRi.002    Iran

QDi.135 - DAWOOD IBRAHIM KASKAR
23 aliases, including Dawood Ebrahim, Sheikh Dawood Hassan, Anis Ibrahim,
  Daud Hasan Shaikh Ibrahim Kaskar and Daud Ibrahim Memon Kaskar
Note: Father's name is Sheikh Ibrahim Ali Kaskar, mother's name is Amina Bi,
  wife's name is Mehjabeen Shaikh. International arrest warrant issued by
  the Government of India.

The same loop in Python

import os, requests

def extract(url, actions=None, driver=None):
    body = {"url": url, "render": True, "formats": ["html"]}
    if actions:
        body["browser_actions"] = actions   # chosen by your LLM from the previous page
    if driver:
        body["driver"] = driver             # "vx10" for JavaScript apps such as the UN search
    r = requests.post("https://sdk.nimbleway.com/v2/extract",
                      headers={"Authorization": f"Bearer {os.environ['NIMBLE_API_KEY']}"},
                      json=body, timeout=300)
    return r.json()["data"]["html"]

page = extract(link)                                   # call 1
for step in range(4):                              # cap the number of steps
    decision = llm_next_step(page, name)               # your model: actions, a new URL, or done
    if decision.done:
        break
    page = extract(decision.url, decision.actions, decision.driver)

Your LLM handles five kinds of page

The Nimble Plugin gives your LLM the same tools

• The plugin installs the Nimble skill and the Nimble MCP server into Claude Code, Cursor or Codex.
• Your engineers can work out the actions for a new site with the LLM before they add the site to your pipeline.
• The plugin's skill drives form searches through the Nimble CLI. The CLI accepts the same browser actions as the API.

npm i -g @nimble-way/nimble-cli
export NIMBLE_API_KEY="your-api-key"

claude plugin marketplace add Nimbleway/agent-skills
claude plugin install nimble@nimble-plugin-marketplace

One agent call returns the financial-crime actions against a subject

• The agent first lists every country where the subject is registered, listed, operates or was investigated.
• It then searches each country's prosecutors, regulators, financial intelligence unit, tax authority, stock exchange and courts by name.
• It searches in the local language and in English, then the US, UK, EU and UN bodies that act on foreign companies.
• It keeps only money laundering, terrorist financing, sanctions, bribery, fraud, tax evasion, regulator enforcement and financial lawsuits.
• Each row says whether the subject was the accused party, so a case where the subject was the victim never reaches your case file.

Example output from your Nimble account

{
  "url": "https://danskebank.com/news-and-insights/news-archive/company-announcements/2025/ca15122025",
  "category": "money_laundering",
  "headline": "Danske Bank confirms conclusion of US Department of Justice probation",
  "language": "English",
  "authority": "U.S. Department of Justice",
  "event_date": "2025-12-13",
  "event_type": "other",
  "publication": "Danske Bank",
  "jurisdiction": "United States",
  "subject_role": "settling_or_fined",
  "authority_page": false,
  "case_reference": "Company Announcement No. 55/2025",
  "additional_urls": null,
  "allegation_summary": "Danske Bank confirmed that the three-year corporate probation imposed under its December 2022 DOJ plea agreement concluded on 13 December 2025.",
  "subject_name_as_published": "Danske Bank"
}

{
  "url": "https://media.bis.gov/media/documents/a781.pdf",
  "category": "sanctions_or_export_control",
  "headline": "BIS fines Boeing over anti-boycott breaches",
  "language": "English",
  "authority": "U.S. Department of Commerce, Bureau of Industry and Security (Office of Antiboycott Compliance)",
  "event_date": "2026-08-06",
  "event_type": "fine_or_penalty",
  "publication": "Bureau of Industry and Security; Global Investigations Review",
  "jurisdiction": "United States",
  "subject_role": "settling_or_fined",
  "authority_page": true,
  "case_reference": null,
  "additional_urls": "https://globalinvestigationsreview.com/just-sanctions/article/bis-fines-boeing-over-anti-boycott-breaches",
  "allegation_summary": "BIS's Office of Antiboycott Compliance charged Boeing with two EAR §760.2(d) violations for furnishing Israeli-boycott certification language on 2019 Dubai Airshow invoices; Global Investigations Review reported BIS fined Boeing.",
  "subject_name_as_published": "The Boeing Company"
}

Send one request per subject

import os, time, requests

API = "https://sdk.nimbleway.com/v2/agents"
H = {"Authorization": f"Bearer {os.environ['NIMBLE_API_KEY']}"}
AGENT_ID = "wsa_..."   # the agent ID we give you when the agent is in your account

prompt = ("Screen Danske Bank A/S. Entity type: company. Home jurisdiction: Denmark. "
          "Known aliases: Danske Bank. Start by listing every jurisdiction where the subject is "
          "incorporated, listed, operates or operated, or has been investigated, then search each one. "
          "Return every in-scope event in any jurisdiction, or an empty findings list if there are none.")

run = requests.post(f"{API}/{AGENT_ID}/runs", headers=H, json={"input": prompt}).json()
while requests.get(f"{API}/{AGENT_ID}/runs/{run['id']}", headers=H).json()["status"] not in ("completed", "failed"):
    time.sleep(30)
result = requests.get(f"{API}/{AGENT_ID}/runs/{run['id']}/result", headers=H).json()
findings = [f for f in result["output"]["content"]["findings"]
            if f["subject_role"] not in ("victim", "witness_or_other")]

curl -X POST https://sdk.nimbleway.com/v2/agents/$AGENT_ID/runs \
  -H "Authorization: Bearer $NIMBLE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "Screen Danske Bank A/S. Entity type: company. Home jurisdiction: Denmark. Known aliases: Danske Bank. Start by listing every jurisdiction where the subject is incorporated, listed, operates or operated, or has been investigated, then search each one. Return every in-scope event in any jurisdiction, or an empty findings list if there are none."}'

# then poll GET /v2/agents/$AGENT_ID/runs/{run_id} and read GET .../runs/{run_id}/result

Each finding is one row with eleven fields

The response also lists every country the agent searched and the authorities it searched in each one.

The agent finds every person with the name and gives each one a verdict

• The agent first runs a LinkedIn people search for the name, then for the name with the place you supply.
• It reads sanctions lists, company registries, regulators and courts, then profiles and press.
• It records each value exactly as printed, with the URL and the name that page prints.
• It then compares each person with the identifiers you supplied and gives one verdict per person.

A strong identifier is the only way to reach same person

The name alone never decides a verdict. Each verdict names the attribute that decided it and the page it was read on.

Example output from your Nimble account

{
  "attestations": [
    {
      "side": "supplied",
      "attribute": "full_name",
      "qualifier": null,
      "person_ref": "P0",
      "source_url": null,
      "source_date": null,
      "source_class": null,
      "value_verbatim": "Gilbert Verdian",
      "about_name_as_printed": "Gilbert Verdian"
    },
    {
      "side": "found",
      "attribute": "full_name",
      "qualifier": null,
      "person_ref": "P1",
      "source_url": "https://find-and-update.company-information.service.gov.uk/officers/quDo4Xesd_el_0RCx9xJ36g157g/appointments",
      "source_date": "2026-10-12",
      "source_class": "corporate_registry",
      "value_verbatim": "Gilbert VERDIAN",
      "about_name_as_printed": "Gilbert VERDIAN"
    }
  ],
  "candidate_verdicts": [
    {
      "reason": "Only the name was supplied; found record shows January 1978, British, Quant director, but no supplied identifier exists to compare.",
      "verdict": "unresolved",
      "person_ref": "P1",
      "found_value": null,
      "supplied_value": null,
      "deciding_attribute": "none",
      "deciding_source_url": null,
      "strong_identifier_match": false
    }
  ]
}

{
  "attestations": [
    {
      "side": "supplied",
      "attribute": "full_name",
      "qualifier": null,
      "person_ref": "P0",
      "source_url": null,
      "source_date": null,
      "source_class": null,
      "value_verbatim": "Brad Jackson",
      "about_name_as_printed": "Brad Jackson"
    },
    {
      "side": "supplied",
      "attribute": "residence_country",
      "qualifier": null,
      "person_ref": "P0",
      "source_url": null,
      "source_date": null,
      "source_class": null,
      "value_verbatim": "Ontario, Canada",
      "about_name_as_printed": "Brad Jackson"
    }
  ],
  "candidate_verdicts": [
    {
      "reason": "Residence Greater Toronto Area, Ontario matches; no conflicting identifier found",
      "verdict": "consistent_unconfirmed",
      "person_ref": "P1",
      "found_value": "Greater Toronto Area, Canada",
      "supplied_value": "Ontario, Canada",
      "deciding_attribute": "residence_country",
      "deciding_source_url": "https://www.linkedin.com/in/brad-jackson/",
      "strong_identifier_match": false
    }
  ]
}

Results on public cases

Send the name and what you already hold

run = requests.post(f"{API}/{AGENT_ID}/runs", headers=H, json={"input": (
    "Sergei Ivanov. Supplied by the investigator: born 31 January 1953, Russian national. "
    "Find every person with this name and judge each one against the supplied details."
)}).json()

# result["output"]["content"] holds:
#   attestations        one row per value found, with its URL and the person it describes
#   candidate_verdicts  one verdict per person, with the deciding attribute and page
#   identity_anchor     the official record behind a same_person verdict
#   collection_gaps     the registries and sites the agent could not reach