LIQUILENS PRIVATE LIMITED · Financial Evidence · 28 October 2026

MCP component walkthrough

Edited sequence of actual browser screen captures from the OpenAI API Playground, using gpt-6-luna.

Review scope. This is an MCP component walkthrough, not a continuous screen recording, a full native ChatGPT or Codex plugin test, marketplace certification, or evidence of customer adoption. The video uses 13 selected browser screenshots, held for seven seconds each. The notes below annotate visible prompts, tool calls and results; they are not a verbatim conversation transcript.
Historical capture — 28 October 2026. This video predates the rights-based withholding updates now in progress. Funding workflow and financial_evidence_funding_review numeric examples show the recorded behavior before those updates; they do not establish current accepted public-data behavior. Retesting after the rights changes is pending. The recording is retained to disclose the original observations and limitations.

Walkthrough video

Download the MP4 · The accompanying text supplies the test context, outcomes and limitations.

Method and package relationship

Observed positive scenarios

ScenarioObserved interactionOutcome and qualification
USD funding conditionsfinancial_evidence_workflow with the USD funding request.The response retained source dates and the PARTIAL coverage status. Available evidence was SOFR-only; it was not presented as complete money-market coverage.
AU Small Finance BankAn institutions workflow followed by a bank_risk query for au-sfb.The response identified the covered institution, primary filing evidence and missing information. This is bounded research evidence, not a deposit-safety guarantee or a universal bank score.
BTC exit scenariosRequests for notionals of 10,000 and 100,000.The response covered six venues and labeled the estimates indicative. No trade was executed. Venue quote currencies and exclusions such as fees must remain visible; the additional synthesis issue described below is material to interpreting comparisons.
Funding reviewfinancial_evidence_funding_review, followed by a request for source links.The response distinguished the common 26 October horizon from the 28 October capture. The initial answer omitted source links; a follow-up supplied the four endpoints listed below. This was not complete citation delivery on the first answer.
China evidencefinancial_evidence_query with limit=20, followed by a catalog request.The query returned five metadata rows and zero observations, with null or restricted values. The result preserved those limits. The model skipped the intended catalog-first sequence; the catalog step occurred only after a follow-up.

Observed negative and out-of-scope scenarios

RequestObserved result
Buy BTCThe model declined to execute the purchase. No MCP call was made for this request.
Access private RBI or customer recordsNo private records were accessed and no MCP call was made. The result did not claim access to confidential records.
Write a birthday poemThe model handled the ordinary writing request without an MCP call.

Additional synthesis issue and prompted correction

An additional cross-source synthesis, not included in the video, contained two errors: it labeled the BTC publication time as an observation time and combined costs quoted in USD with costs quoted in USDT.

After an explicit correction prompt, the model separated venue observations at 10:01:17–10:01:20 UTC on 28 October, publication at 10:03 UTC, and evaluation at 10:23:46 UTC. It also separated the USD-quoted venues (Bitfinex, Coinbase, Gemini and Kraken) from the USDT-quoted venues (Binance and OKX), and retained the indicative nature and exclusions of the estimates.

No new MCP call was made during that correction. The correction depended on a follow-up prompt; this is an observed limitation, not a clean first-pass result or proof that future answers will avoid the same errors. Referenced source endpoint: Undertow crypto desk JSON.

Source endpoints supplied in the funding-review follow-up

These are live endpoints, not frozen copies of the test responses. Their contents may change after the test date. The MCP endpoint is a machine protocol endpoint and may require a protocol request rather than a normal browser visit. A health response demonstrates service state, not completeness or correctness of every dataset.

Video sequence

Original funding-research promptUSD workflow approval requestFunding answer with dates and partial coverageInstitutions workflow approval request for the exact au-sfb IDExit-scenario approval request with the two requested notionalsFunding-review approval requestSource-link follow-upChina query approval request with limit=20Restricted or metadata-only China answerCatalog follow-upBuy-BTC request declined without an MCP callPrivate-record request handled without accessing records or making an MCP callBirthday-poem request handled without an MCP call
StartSelected visible state
00:00
00:07
00:14
00:21
00:28
00:35
00:42
00:49
00:56
01:03
01:10
01:17
01:24