TCAML — Open AML Agent Framework

TCAML is an open-source framework for building local, human-supervised AML/KYC/KYB agent workflows.

It combines AML skills, operation workflows, evidence-control rules, structured schemas, a deterministic local runner, and an LLM orchestration interface. The goal is simple: help local AML agents draft review-ready case files while keeping gate confirmations and final compliance decisions under human control.

Why this exists

AML work is not just search. It is a controlled process of checking declared information, collecting evidence, recording discrepancies, identifying risk triggers, and handing the case to the responsible reviewer.

TCAML gives developers and compliance teams a public foundation for building that process into local agents, RAG systems, coding assistants or internal AML operations tools.

What you can build with it

• Client onboarding review workflows
• Counterparty and supplier screening workflows
• KYB and UBO review support
• Employee screening workflows
• Periodic review and rescreening workflows
• EDD escalation and MLRO handover support
• Evidence packs, discrepancy registers and audit trails
• Local AML agent prototypes using structured LLM output

How it works

AML skills and operation workflows
→ S010 verification and evidence control
→ LLM content draft, if used
→ validated S010 JSON
→ deterministic local runner
→ human-confirmed stagegates
→ case folder, reports and audit trail
→ Pending human review

The LLM may draft structured content. It must not approve cases, confirm gates or make final compliance decisions.

Main components

skills/                  AML skills, operation workflows and source guidance
schemas/                 Structured output schemas and gate keyword registry
templates/               Evidence, review and case-file templates
tcaml_runner/            Optional local runner with exact-keyword stagegates
tcaml_orchestrator/      Optional provider interface for LLM content drafting
tools/                   Validation, JSON parsing and case-folder utilities
tcaml-agent-framework/   Prompts, retrieval routing, output contracts and safety boundaries
implementation-profiles/ Generic public implementation profile example

The word skills remains important: the framework is built from reusable AML/KYC/KYB skills that can be loaded by agents and combined into operation workflows.

Try the local demo

Install dependencies:

pip install -r requirements.txt

Run tests:

python -m pytest -q

Create a validated S010 draft with the mock orchestrator:

python -m tcaml_orchestrator draft --provider mock

Run the local human-gated workflow:

python -m tcaml_runner run dist/orchestrator/s010-content.json

Generated demo files are written under dist/.

Local LLM support

TCAML includes an OpenAI-compatible provider interface for local model servers that expose /v1/chat/completions, including LM Studio, vLLM, llama.cpp server configurations and compatible gateways.

Example:

export TCAML_OPENAI_BASE_URL=http://localhost:1234/v1
export TCAML_OPENAI_MODEL=local-content-drafting-model
python -m tcaml_orchestrator draft --provider openai-compatible

The orchestrator validates model output through the hardened LLM-to-S010 bridge. Normal JSON or schema errors can produce correction prompts. Safety violations stop the process.

Human review boundary

TCAML is not an automated approve/reject engine.

For S010 workflows, the agent-produced decision boundary is always:

Pending human review

Gate confirmations must be entered by a human analyst. LLM-generated drafts cannot trigger --auto-confirm and cannot control gate status, gate event logs or final human-review decisions.

What this repository does not include

This repository does not include AML database access, paid source credentials, customer data, live evidence packs, internal risk thresholds, legal advice or final compliance decisions.

Organizations using TCAML must configure their own source access, policies, thresholds, permissions and review responsibilities.

Start here

• skills/operations/ — AML/KYC/KYB operation workflows
• skills/core/TCAML-S010-verification-evidence-control/ — verification, evidence control and stagegates
• schemas/tcaml-s010-output.schema.json — main S010 output schema
• tcaml_runner/ — local deterministic runner
• tcaml_orchestrator/ — model-agnostic LLM drafting interface
• docs/public-vs-private-boundary.md — public/private implementation boundary

License

MIT License. See LICENSE.

Maintained by Tom Custos

TCAML is maintained by Tom Custos as an open AML agent framework for local, evidence-based AML/KYC/KYB workflows.

For Commercial Implementation Profiles, local AML agent deployment, source workflow configuration, evidence rules, MLRO escalation logic or commercial support, contact at hello@tomcustos.com.