Skip to main content
W-2 + HITL available · full LOS shipping with vertical package

Agent-native lending and mortgage operations.

ClawQL’s lending vertical (clawql-lending) composes the IDP pipeline, vault memory, HITL review, and workflow automation for loan origination — W-2s, bank statements, and underwriting packages without pasting specs or losing audit trails.

Lending compose stacks and W-2 samples use synthetic data only — not tax, legal, or underwriting advice. Train tenant-specific classifiers before production.
Overview

ClawQL for lending

Mortgage, auto, BNPL, payday, and commercial lending share the same document-heavy intake problem: borrowers submit heterogeneous PDFs, underwriters re-ask for the same conditions, and API integrations sprawl across LOS, credit, and document vendors. Modularization v2.1 defines clawql-lending as five sub-verticals on one Agentic Gateway — MortgagePlugin, AutoPlugin, BNPLPlugin, PaydayPlugin, and CommercialPlugin — plus shared UnderwritingPlugin and CompliancePlugin for Reg Z, ECOA, and fair-lending guardrails. Today you can run the lending Docker Compose stack and the shipped W-2 reference pack; full LOS tool registration ships with the vertical package.

ClawQL powers vertical lending products like See The Greens LOS — a mortgage-first loan origination system built on the same IDP, vault, and HITL stack documented here.

Challenges

Problems agents solve in this vertical

These are the operational friction points clawql-lending targets — on top of ClawQL Core search, execute, memory, audit, and the IDP pipeline.

Spec dumps burn planning tokens

Credit bureau, GSE, and LOS OpenAPI surfaces are enormous. Pasting them into agent prompts is expensive and brittle. search returns ranked operationIds; execute validates args server-side.

Context dies between underwriting sessions

Cleared conditions and vendor analysis live in email threads. memory_ingest and memory_recall give underwriters durable borrower context across Cursor, OpenClaw, and your LOS-adjacent automations.

Low-confidence OCR needs humans in the loop

W-2 box misreads create repurchase risk. classify_document routes by type; hitl_enqueue_label_studio suspends Argo workflows until a reviewer confirms fields.

Platform

Shared ClawQL capabilities

Every vertical package composes these horizontal layers — security, memory, and document intelligence.

  • Eight-vendor IDP pipeline (Docling, Tika, Gotenberg, Stirling, archive, Onyx, Coneshare)
  • Vault memory with wikilinks for borrower files and underwriting decisions
  • Argo Workflows suspend/resume integrated with Label Studio HITL
  • Cosign-signed MCP images with Kyverno admission on Helm deploys
  • Optional Fabric consortium features via clawql-blockchain peer dependency
Domain tools

Tools from clawql-lending

Planned or shipping MCP tools from modularization v2.1 — registered when the vertical package is enabled via CLAWQL_ENABLE_* or Operator flags.

gse_validate()

Validate loan data against agency guidelines indexed in Onyx.

condition_clear()

Record and recall underwriting conditions with vault threading.

pii_redact()

Strip borrower PII before documents enter search indexes.

loan_archive()

Tag and archive loan packages in the ClawQL Archive Layer or Paperless-ngx on self-hosted stacks.

bnpl_decision / fraud_check()

Sub-second decision loops for BNPL with Cuckoo dedup on doc batches.

Document types

What the IDP pipeline processes

Representative inputs agents classify, extract, redact, and archive in this vertical.

  • W-2 and 1099 income forms
  • Pay stubs and VOE letters
  • Bank statements and asset statements
  • Appraisal and title commitments
  • Purchase agreements and closing disclosures
Use cases

Why ClawQL for lending

Vertical packages extend the same Agentic Gateway — search, execute, memory, IDP, audit — with domain tools from clawql-lending.

Income document ingestion

Parse W-2s and pay stubs with Docling layout OCR, classify by document type, route low-confidence extractions to Label Studio, and archive tagged outcomes for underwriters with correspondent metadata.

Underwriting memory across sessions

memory_recall surfaces prior borrower conditions, cleared stipulations, and vendor analysis — underwriters do not re-derive context every time a loan file reopens in a new agent thread.

Multi-product LOS coverage

Mortgage, auto, BNPL, payday, and commercial modules share compliance plugins but register distinct MCP tools — enable only the products you originate via CLAWQL_ENABLE_* and Operator vertical flags.

Compliance-aware automation

Shared CompliancePlugin targets Reg Z, ECOA, and fair-lending checks. audit logs structured MCP events; dedicated managed hosting isolates tenant workloads for strict lenders.

Examples

Example agent workflows

Representative MCP tool sequences with step-by-step detail — production deployments add tenant classifiers, RBAC, and your compliance policies on top.

Shipped reference path — synthetic W-2 through Docling, classify, HITL, vault.

W-2 intake with confidence routing

The lending W-2 pack demonstrates the full pattern agents use for income verification: parse, score confidence, human review when needed, persist decisions durably.

  1. 1Parse layoutexecute docling_convert_file on fixtures/synthetic-w2.txt — layout-aware OCR for boxed forms.
  2. 2Classifyclassify_document routes W-2 vs pay stub vs bank statement before extraction schema selection.
  3. 3Route low confidenceScore below threshold → hitl_enqueue_label_studio with Label Studio project from label-studio-config.xml.
  4. 4Persist reviewWebhook fires → memory_ingest appends reviewer fields with wikilinks to the loan file note.
execute()classify_document()hitl_enqueue_label_studio()memory_ingest()

Production-style workflow template with explicit suspend step.

Argo suspend / resume for underwriting HITL

workflow submits clawql-lending-w2-ingest; operators resume after Label Studio without losing pipeline state.

  1. 1Submit workflowworkflow operation submits WorkflowTemplate clawql-lending-w2-ingest with confidence_threshold parameter.
  2. 2SuspendTemplate hits suspend when classify confidence < threshold — file waits in Argo, not in chat context.
  3. 3Human reviewUnderwriter completes Label Studio task; webhook validates CLAWQL_HITL_WEBHOOK_TOKEN.
  4. 4Resumeworkflow resume continues to finalize step; audit captures correlation_id for the loan.
workflow()hitl_enqueue_label_studio()memory_recall()

OpenClaw or Cursor picks up where the last underwriter left off.

Cross-session borrower recall

A fresh agent thread calls memory_recall with the borrower name and condition keywords — vault returns ranked snippets without re-uploading prior PDFs.

  1. 1Recallmemory_recall query "borrower Smith conditions cleared 2026" with maxDepth for wikilink graph.
  2. 2Discover APIssearch finds LOS or credit APIs for status updates — no 9 MB spec in prompt.
  3. 3Executeexecute updates condition status with validated args and bearer token from env.
  4. 4Ingest outcomememory_ingest appends what changed with append: true on stable title.
memory_recall()audit()search()

POC underwriting environment via Docker Compose.

One-command local lending stack

docker compose -f lending.compose.yml brings up MCP, Docling, classifier, LangExtract demo, and Label Studio CE — documented in docker/compose/README.md.

  1. 1Bootstrapcp lending.env.example; docker compose up -d --build.
  2. 2Wire HITLPaste label-studio-config.xml; set CLAWQL_LABEL_STUDIO_API_TOKEN; configure webhook to /hitl/label-studio/webhook.
  3. 3Connect agentPoint Cursor or OpenClaw at http://localhost:8080/mcp (streamable-http).
  4. 4Smoke W-2 flowRun classify → HITL path from deployment/samples/lending-w2/README.md.
search()execute()audit()
Security & compliance

Built for regulated workflows

Industry pages summarize platform capabilities — your legal, compliance, and security teams should review docs.clawql.com/security before production data.

  • PII redaction via Stirling in the IDP pipeline before archive and Onyx indexing
  • Structured audit ring buffer plus optional vault postmortems for examiner requests
  • Vertical RLS and ATR scoping planned via clawql-auth — underwriter role bound to mortgage vertical
  • Cosign-signed images and Kyverno verifyImages on self-hosted Helm
  • Fair-lending and Reg Z hooks in shared CompliancePlugin (modularization v2.1)

Ready to demo ClawQL for lending?

Self-host the lending compose stack today, or join the managed waitlist for hosted MCP, vault, and IDP tailored to lending workflows. Start your 14-day trial or self-host free with npm or Helm — full Apache 2.0 stack, no license fee.