LiquiLens

LiquiLens MCP — the Failure Radar as agent tools

Endpoint: https://api.liquilens.in/mcp (streamable HTTP, no auth, no install)

Try it live: liquilens.in/developers · REST catalog: api.liquilens.in/api

LiquiLens is a failure early-warning system for banks and lenders, built on public data with a machine-readable historical-evidence boundary served beside every claim. This MCP 1.8.0 endpoint exposes 21 read-only tools and 5 guided prompts so an agent reads the same status, eligibility flags and cited record a human sees. Its capability inventory is pinned to LiquiLens commit ad759890fbd3656397c3c28ab8f3b0b845aeede9.

It now reads both ends of the chain: which institutions are fragile, and whether that stress is actually crossing into the real economy: companies rolling paper and drawing credit lines, households falling behind. Channels that cannot be read are named in cannot_see rather than reported as calm.

Add it

Claude Code:

claude mcp add --transport http liquilens https://api.liquilens.in/mcp

Codex:

codex mcp add liquilens --url https://api.liquilens.in/mcp

Cursor: merge the remote MCP config into .cursor/mcp.json for one project or ~/.cursor/mcp.json for all projects. The download includes the companion Seiche funding endpoint. Preserve existing entries. For an existing Codex config, use the TOML snippet and merge its server entries into ~/.codex/config.toml.

Claude.ai / ChatGPT: add a custom connector or MCP app with the endpoint above where that feature is available in your workspace.

This repository is the public discovery and documentation mirror. The hosted implementation is maintained in a private core repository. This public mirror pins its inspectable capability contract to the reviewed source revision above. The official Registry already serves io.github.beepboop2025/liquilens version 1.8.0. This update reconciles the mirror with that release; it does not publish a new Registry version.

Run a research task

The research recipes run with Python 3.11+ and its standard library. No API key, LLM or Python package install is needed. Download and inspect financial_research.py, then run:

python3 financial_research.py bank-review
python3 financial_research.py bank-review --slug cosmos-ucb
python3 financial_research.py funding-brief

The first command discovers covered bank slugs. The second checks coverage before retrieving that exact bank's sourced asset-quality history; an absent slug stays not_covered. The funding brief calls Seiche's money-market desk. Returned evidence retains its native source dates, eligibility, unavailable states and limitations; a successful request does not imply fresh or complete evidence. Each run is bounded and performs no scheduling or trading. Operators should add --verification so their checks are labelled synthetic and excluded from adoption totals.

For a visual workflow, download the n8n bank-review workflow and follow the import guide. It uses a manual trigger and no LLM. The guide records its execution-verification status; a downloadable workflow does not imply acceptance into n8n's template library.

Protocol compatibility

• 2026-07-28: stateless requests use server/discover, per-request _meta, MCP-Protocol-Version, and mirrored Mcp-Method / Mcp-Name routing headers.
• 2025-11-25, 2025-06-18, and 2025-03-26: retained legacy initialization, tools, prompts, notifications, batching, and ping behavior.
• Tools, prompts, and discovery responses are deterministic. Modern list responses carry public cache metadata.
• resources/list and resources/templates/list are supported and currently return empty catalogs. resources/read returns an explicit not-found error instead of inventing a resource.

Tools

Tool	What it serves
bank_asset_quality_review	Exact-slug bank review with sourced GNPA/NNPA history, percentage-point changes, distinct PCR definitions and capital/supervisory scope; no new score or credit approval
bank_npa_reconciliation	Arithmetic check of a complete caller-supplied NPA movement table; keeps cash recoveries, write-offs, sales and upgrades distinct without authenticating the input
banking_specialisation_coverage	Discover covered Indian commercial, small finance and urban cooperative banks, with observed, stale, historical and absent evidence distinguished; not a census or rating
corporate_transmission_board	Is funding stress reaching nonfinancial firms? The commercial-paper market, bank credit lines, real-economy confirmation (claims, capex, inventories, openings, business bankruptcies) and a balance-sheet context channel, with a TRANSMITTING/CONTAINED verdict
crypto_exposure_board	Cited bank/crypto exposure register joined to Undertow run-risk context; display-only, never an institution score
crypto_regime_board	Compact BTC/ETH change-point state and its display-only cross-read against disclosed bank exposure
evidence_europe	NAMED_CASE_FILES_CONSTRUCTION_PIT: seven audited case files, deliberately no cohort claim; both eligibility flags false
evidence_india	PERIOD_END_PROXY_CONSTRUCTION_PIT: 48 institutions across two decades, misses and false alarms included; both eligibility flags false
evidence_institution	One Indian institution's construction-PIT crisis replay with sourced quarterly rows
evidence_markets	Historical-evidence status for all three markets, with validated_backtest_eligible and real_money_eligible served explicitly
evidence_us	CURRENT_AMENDED_CONSTRUCTION_PIT: 552 FDIC failures since 2008, 72.8% recall, 21.7-month median lead and AUC 0.854; both eligibility flags false
failure_radar_board	The live board: every Indian lender with a fresh vetted dossier, including failure PD term structure (12/24/36m), RBI action-zone status, funding fragility, market distance to default, and watchlist tier under a published rule
failure_radar_institution	One institution in depth: PD trajectory with drivers, PCA/SAF headroom history, forensic screen, market reading
forward_odds	Counted forward stress odds for each public-signal layer, withheld until the layer has enough observed history
household_credit_board	Is stress transmitting through household balance sheets? Fed delinquency and charge-off legs against each leg's own trailing decade, two-sided revolving-credit velocity, debt service as unscored context
institution_review_packet	One lender's deterministic evidence packet for human review, with coverage and freshness stated explicitly
latest_article	Today's exact evidence-led LiquiLens article, or a labelled historical replay when the evidence did not move
rbi_supervisory_tape	Latest RBI enforcement actions, each linking to the RBI's own page
stablecoin_rails_board	Issuer peg, redemption-run, chain-concentration and rail tripwire state; missing data never becomes CALM
universe_search	RBI's official registered-NBFC registry (9,000+ entries)
verify_published_record	Independent cryptographic verification of the as-published record

Prompts

Prompt	Guided playbook
bank_asset_quality_brief	Discover an exact bank slug, review cited NPA history and capital scope, then reconcile only a complete disclosed movement table; preserve stale/missing evidence and supervisory gaps
crypto_liquidity_briefing	BTC/ETH regime, stablecoin rails, and disclosed bank links in one display-only evidence pack
failure_radar_briefing	Board-level institution risk with transmission context from the public US signal layers
institution_health_check	One lender health check with the historical-evidence boundary and uncertainty attached
stress_evidence_pack	Region-specific evidence for India, the United States, Europe, or the cross-market summary

The governance line

The generative layer explains; deterministic screening policy scores. Historical diagnostics retain their served evidence status and eligibility flags and do not become validated-backtest or real-money evidence merely because a deterministic engine produced them. Screens, not ratings; not investment advice. Institutions without a vetted dossier are absent by design — the tools say so rather than inventing a score.

The three historical status tokens are part of the API contract, not marketing copy: PERIOD_END_PROXY_CONSTRUCTION_PIT for India, CURRENT_AMENDED_CONSTRUCTION_PIT for the United States, and NAMED_CASE_FILES_CONSTRUCTION_PIT for Europe. In the current release, every market serves validated_backtest_eligible: false and real_money_eligible: false.

Siblings from the same lab

• Seiche — US money-market funding stress (the plumbing)
• groundcheck — claim grounding and citation verification
• Palimpsest (https://api.seiche.info/palimpsest/mcp) — live internet-censorship signals
• Undertow — the cross market liquidity map: daily tiered board, exit cost at position size, Telegram front door at t.me/undertow_LiquiLens_bot

Product: liquilens.in · Live demo: demo.liquilens.in