HomeArchitectureDocsGlossaryAPI referenceStatus

CAIN TRUST FABRIC · architecture

One platform. Two deployment models.

AI trust infrastructure for autonomous systems. CAIN Trust Fabric is the parent platform; CAIN Studio is its hosted deployment and MCPGate is its self-hosted one. Same architecture, same domains, different operator.

AI can now act. CAIN Trust Fabric makes those actions provable, governed, and enforceable.

CAIN TRUST FABRIC · in ten seconds

AI agents can now act. We make every action provable, governed, and enforceable.

CAIN Trust Fabric sits between your agents and the things they do. Before an action runs, it checks who is asking, whether policy allows it, how risky it looks, and whether the plan can break your rules -- then it enforces that decision and writes down the evidence.

One identity for every actorAgents, humans and services in one namespace, with per-agent budgets you can revoke instantly.
A decision before every actionIdentity, policy, risk and formal plan-verification run as one call and return one verdict.
Enforcement in the call pathNot a dashboard that tells you afterwards -- a boundary that can stop the call, hosted by us or self-hosted by you.
Evidence you can produce laterEvery decision stored with its stage-by-stage verdicts and retrievable by id.

The eleven trust domains: Identity · Policy · Authorization · Risk · Governance · Verification · ActionProof · Security · Execution Enforcement · Evidence · Observability.

Works with what you already use. Works with any agent runtime -- LangChain, LangGraph, LlamaIndex, CrewAI, OpenAI, Google, Microsoft, AWS or your own loop -- because it's reached over plain HTTPS and MCP. No rewrite, no SDK lock-in.

Two ways to run it. You're on CAIN Studio, the hosted deployment; MCPGate is the other one. Two ways to run it: CAIN Studio (hosted, subscription, nothing to install) or MCPGate (self-hosted, perpetual license, your network and your evidence). Either one unlocks the other.

What we don't claim. Enforcement starts in shadow mode -- every check runs and is recorded, nothing is blocked -- until you've seen what it would have stopped on your own traffic. No certifications are claimed.

See the full architecture → Get it as JSON → Browse every service →

The decision path

Five steps, in the order they actually execute. Every gated call on this deployment produces one decision record with the stage-by-stage verdicts that produced it.

1
THE AGENT PROPOSES.
Any runtime -- LangChain, LangGraph, LlamaIndex, CrewAI, an OpenAI/Google/Microsoft/AWS agent stack, or your own loop -- submits an intended action or plan over HTTP or MCP.
2
CAIN DECIDES.
The control plane resolves the caller's identity, evaluates policy, scores risk, and asks the verification layer whether the plan can violate the tenant's constraints. One decision record, one verdict.
3
MCPGATE ENFORCES.
The decision lands on an enforcement boundary that sits in the call path -- hosted at the CAIN Studio gateway, or self-hosted inside your own network as MCPGate.
4
THE SYSTEM EXECUTES.
Authorized calls proceed to the real upstream service. Blocked calls never reach it, and the caller is told which stage stopped them and why.
5
EVIDENCE REMEMBERS.
Every decision is written to a durable, tenant-scoped decision record with the stage-by-stage verdicts that produced it, retrievable by id.

Deployment models

Hosted / Managed CAIN Trust Fabric

CAIN Studio You are here

The managed control plane. We run it; you point your agents at it.

Enforcement boundary
Hosted enforcement boundary at the CAIN Studio platform gateway.
Commercial model
Subscription (Free, Professional and Enterprise tiers), per-call metering on metered services.
Who it's for
Teams that want the Fabric working today without operating it.
Control plane
Operated by us
Self-Hosted / Customer-Controlled CAIN Trust Fabric

MCPGate Visit MCPGate →

The enforcement boundary you run yourself, inside your own network, where the traffic and the evidence never leave your infrastructure.

Enforcement boundary
Customer-operated enforcement boundary in the customer's own network.
Commercial model
One-time perpetual license (bundle or per-tool), self-hosted, no per-call metering by us.
Who it's for
Teams whose data, or whose regulator, will not accept a hosted control plane.
Control plane
Operated by you

A subscription on CAIN Studio and a license on MCPGate unlock each other through the cross-platform entitlement bridge, so choosing a deployment model is not a choice about which products you get.

The eleven Trust Fabric Services

The trust core. These are the services that decide, enforce, and remember -- the eleven domains the Fabric is made of. Counts below are computed from this deployment's live routing table, so a service that isn't deployed can't be counted.

Identity

5 services here

One namespace for every principal in the loop -- agents, humans, and services.

Know exactly which agent, person, or service is making this call, and prove it.

agent-idcainaccountsa2a-guardtaloscrypt-keymgmttaloscrypt-pqc

Policy

6 services here

The rules an action has to satisfy, authored per tenant and evaluated per call.

Write down what is and isn't allowed -- including the regulation you're actually subject to.

caingoverncompliance-enginelegal-auditorquorum-i18ntalosdoc-contracttaloslegal-caselaw

Authorization

3 services here

Whether this principal may take this action, now, within this budget.

Decide per-call: is this agent allowed to spend this, and does a human need to approve it?

cainpaycainbudgethuman-in-loop

Risk

10 services here

How dangerous this input, output, agent, or fleet looks right now.

Score it before you trust it -- drift, bias, hallucination risk, contradiction, anomaly.

sentinelbias-detectorcaindrifthallucination-firewallcontext-healthmemory-integritycontamination-scannertalosdata-qualitymetacognitive-enhancertrust-state-engine

Governance

4 services here

Who decided, under what mandate, with what dissent recorded.

Show a regulator or a board that the decision had an owner and a paper trail.

blizzard-governancetalosops-governanceguardian-scorecardlexisguardian

Verification

11 services here

Proof that an output, claim, or artifact is correct before anything trusts it.

Check the answer, the citation, the diff, or the derivation -- deterministically, not with another LLM.

rag-verifierveritasenginederivation-copilottaloslogic-prooftalos-verifyvericoding-gatecainschemacruciblecainbenchtopological-data-analysistrajectorygate

ActionProof

2 services here

Formal proof that a planned action sequence cannot violate the tenant's constraints -- before it runs.

Prove the plan is safe ahead of execution, or get the counterexample that shows it isn't.

actionproofaction-firewall

Security

13 services here

Keeping attackers, injections, and leaks out of the agent loop.

Stop prompt injection, jailbreaks, secret leakage, and vulnerable code reaching production.

adversarial-fuzzertalos-redteampii-redactorshadow-agent-scannertaloscode-securitytaloscrypt-audittaloscrypt-vaulttaloscrypt-storetaloscrypt-meshtalosguard-inteltalosguard-siemtalosguard-logiqsecops-patcher

Execution Enforcement

2 services here

The boundary in the call path that actually stops or shapes the call.

Make the decision binding -- not advisory -- at the moment of execution.

talos-shieldswarm-orchestrator

Platform machinery in this domain, not catalog services:

Evidence

7 services here

A durable record of what was decided and what happened, retrievable after the fact.

Answer 'prove this action was authorized' months later, with the original verdicts attached.

content-provenancecainwitnesszkproofvaultmulti-agent-failure-traceragent-debuggerattestationdecision-intelligence

Observability

3 services here

What the system is doing right now, per service and per fleet.

See health, cost, latency, and behaviour while it's happening -- not in a postmortem.

observabilitytalosops-selfheallatent-inspector

Governed workloads

Products that run on the Fabric and are governed by it, but are not part of the trust core: codegen, ML infrastructure, memory, reasoning, and industry applications. Listed separately on purpose -- calling a credit underwriter a trust service would inflate the fabric with something that isn't one.

GroupServices
Build & ship code
Write, review, migrate, and refactor the code itself.
talos-codertalos-evolvetaloscode-reviewtaloscode-migratetalosdev-refactortalosdev-archdiffvisual-ide
Model & ML infrastructure
Quantize, compress, merge, compile, route, and train.
talosai-quantizemodel-compressorweight-alchemistedge-optimizersparselogic-compilersynthetic-fine-tunertraining-sandboxomegarouter
Memory, retrieval & knowledge
Persistent memory, real retrieval, and a shared knowledge store.
memorymeshtalos-ragtalos-vectorsearchknowledge-graphnexusmind
Reasoning, planning & orchestration
Causal inference, search, physical/scene reasoning, and multi-step execution.
causal-networkmcts-enginestcnautonomous-task-enginespatial-synthesizerechoworkforcecausal-intelligencecounterfactual-intelligence
Commerce & distribution
Billing models and marketplaces for AI products.
quorum-oaasagent-marketplace
Industry applications
Domain products built on the same platform: finance, energy, logistics, credit.
cainloancaingridcainroutetalosfin-statarbtalosfin-earningsaiarbitrage
Platform operations
Running the platform itself: experiment analysis and failure intake.
feedback-engineexperimentation-notebook

Which agent runtime you use is not our business

Runtime-agnostic by interface, not by integration deal.

The Fabric is reached over plain HTTPS and over MCP (streamable HTTP). Anything that can make an HTTP request or speak MCP calls the same endpoints with the same headers: LangChain, LangGraph, LlamaIndex, CrewAI, OpenAI, Google, Microsoft and AWS agent runtimes, or a custom loop you wrote yourself. We ship a Python SDK and an MCP server; everything else is one HTTP call. There are no vendor partnerships behind this list and none are implied -- it works because the boundary is a protocol, not a plugin.

LangChainLangGraphLlamaIndexCrewAIOpenAI AgentsGoogle (Vertex / ADK)Microsoft (AutoGen / Azure AI)AWS (Bedrock Agents)Custom / in-house runtimes

Check this page against the running system

Everything above is rendered from one module on both sites, and every claim in it has an endpoint you can call:

curl https://cainstudio.online/fabric/hierarchy      # this hierarchy, as JSON
curl https://cainstudio.online/fabric/domains        # the eleven domains + live service counts
curl https://cainstudio.online/fabric/status         # which stages are live, and whether they enforce

Scope, stated plainly: nothing here is certified against any standard, and no formal guarantee is offered beyond what each individual service proves on its own. Fabric enforcement is opt-in per deployment — /fabric/status reports whether this deployment is enforcing or observing right now.