llms.txt

# AgenTrust

> Free, open specifications and tools that make a signed receipt for an AI system: which model ran, which agent acted, and each step it took. On supported confidential-computing chips, the processor itself vouches for where the work ran (hardware attestation). Anyone can check a receipt on their own computer, offline.

AgenTrust lives at https://agentrust-io.com, with its source code in the GitHub organization https://github.com/agentrust-io. It covers four steps that can be used separately: the model file (Weight Custody Manifest), the agent's identity (Agent Manifest), each action the agent takes (cMCP for tool calls, cA2A for hand-offs between agents) and the receipt format (TRACE). The start pages and demos run in software mode: the receipts are real, but no chip vouches for them. For exact requirements, current release versions, licenses and setup needs, use each project's own documentation linked below.

## Start here

- [Overview](https://agentrust-io.com/): The verifiable AI supply chain from model weights to agent actions, the hardware it is validated on, and what the evidence does and does not prove.
- [Verify an Intel TDX quote](https://agentrust-io.com/verify/): Runs the four-step DCAP check on a genuine GCP confidential VM quote in the browser, ending at the pinned Intel SGX Root CA, then checks that the quote's REPORTDATA commits to the signing key of a TRACE record published beside it. A pass proves genuine Intel TDX silicon signed the quote and bound that key. The check does not appraise TCB currency or revocation, or show that the measured image is the one anyone intended.
- [Get started](https://agentrust-io.com/quickstart/): One start page per specification, in chain order: Weight Custody Manifest, Agent Manifest, cMCP, cA2A, TRACE and the TRACE conformance suite, plus the in-browser TDX quote check, which needs no install. The linked start pages run in software mode with no hardware isolation.
- [Weight Custody Manifest (WCM)](https://wcm.agentrust-io.com/): Bind model-weight identity and custody terms to key-release policy. The local walkthrough uses synthetic evidence and a placeholder key; it does not load a real model or demonstrate hardware protection.
- [Runnable demos](https://agentrust-io.com/demos/): Software examples for policy decisions, evidence verification, delegation, and model-weight custody. Follow each demo's stated prerequisites and limits.
- [Marketplace search](https://agentrust-io.com/marketplace/): Find integrations by framework, technology, and use case.
- [Complete marketplace catalog](https://agentrust-io.com/marketplace/catalog/): Dated integration descriptions and source links, readable without JavaScript or external catalog requests. Listings do not establish endorsement, certification, or hardware validation.

## Technology and trust boundaries

- [Agent Manifest](https://manifest.agentrust-io.com/): Declares agent identity, capabilities, and intended authority. Verification of a declaration does not prove that every runtime action complied with it.
- [Confidential MCP (cMCP)](https://cmcp.agentrust-io.com/): Evaluates routed tool calls against policy and records the decisions. Hardware deployments place the runtime inside a Trusted Execution Environment; confidentiality also depends on egress policy and deployment configuration. The agent and upstream tool server are separate boundaries.
- [Confidential A2A (cA2A)](https://ca2a.agentrust-io.com/): A developer-preview trust profile for Agent2Agent delegation, with attenuated authority, attestation, sealed channels, and provenance. Assurance depends on the evidence and checks performed by the deployment.
- [TRACE](https://trace.agentrust-io.com/): Portable, signed runtime evidence. A Series of LF Projects, with an AAIF Sandbox proposal open. Hardware provenance requires attestation verification against a trusted root. A signature alone does not establish hardware origin or the truth of all recorded claims.
- [WCM public SDK and demos](https://wcm.agentrust-io.com/): Public evaluation path for weight custody. Key release, encrypted artifact delivery, runtime leases, and key-copy handling are separate responsibilities. The specification is in pre-standardization review; use the public material linked from this page.
- [AgenTrust Telemetry](https://agentrust-io.com/telemetry/): An alpha, backend-neutral event contract and reference SDKs for governance facts, caller-owned OpenTelemetry, and evidence finalization. It is not a policy engine or a proof that recorded events are true.
- [TRACE Registry](https://agentrust-io.com/registry/): Public anchors and signed checkpoints for offline verification. Inclusion, a signer's stated time, independent time evidence, and consistency of log history are different checks. A registry entry does not validate a record's claims.
- [Agentic controls](https://agentrust-io.com/go/): Thirty requirements with stable individual permalinks and OpenCRE mappings. A mapping is not certification or a conformance result.

## Specifications, schemas, and tests

- [TRACE schema](https://agentrust-io.com/schema/trace-v0.2.json): Machine-readable record schema.
- [TRACE conformance suite](https://tests.agentrust-io.com/): Test requirements, error codes, and implementation guidance.
- [cA2A extension identifier](https://agentrust-io.com/extensions/ca2a/v0.1/): Profile discovery and Agent Card declaration.
- [Example cA2A Agent Card](https://agentrust-io.com/examples/ca2a-agent-card.json): Documentation fixture, not a live agent endpoint.
- [GitHub organization](https://github.com/agentrust-io): Source, releases, issues, governance, and contribution instructions.
- [Governance resource directory](https://governance.agentrust-io.com/): Curated tools, standards, and research.

## Research

- [Research reports](https://agentrust-io.com/research/): Technical reports on TRACE, cMCP, Agent Manifest, Weight Custody Manifest, cA2A and related problems, each with explicit limits. Published reports are versioned, with PDFs, source packages, recorded software evidence and citations. Not peer reviewed; use current specifications for implementation.

## Community and stewardship

- [Community, partners and contributors](https://agentrust-io.com/community/#adoption): Published roles and partner relationships, each linked to public evidence. AgenTrust is an independent open-source project. No company funds, owns or directs it; contributions are voluntary. The NSF submission is in review, not an award.
- [Architecting at Scale](https://agentrust-io.com/community/#book): The field guide and its public companion code.

## More documentation

- [TRACE summary](https://trace.agentrust-io.com/llms.txt)
- [Agent Manifest summary](https://manifest.agentrust-io.com/llms.txt)
- [cMCP summary](https://cmcp.agentrust-io.com/llms.txt)
- [cA2A summary](https://ca2a.agentrust-io.com/llms.txt)
- [Weight Custody Manifest summary](https://wcm.agentrust-io.com/llms.txt)
- [TRACE conformance suite summary](https://tests.agentrust-io.com/llms.txt)
- [Governance resource directory summary](https://governance.agentrust-io.com/llms.txt)
- [Sitemap](https://agentrust-io.com/sitemap.xml): Canonical public pages on this host.