Agent Route Origin Authorization (AgentROA): A Cryptographic Policy Enforcement Framework for AI Agent Actions over the Model Context Protocol (MCP)

1.1. Motivation

Autonomous AI agents are increasingly being deployed in enterprise and regulated environments. These systems can investigate incidents across cloud infrastructure, invoke external tools, initiate transactions, and delegate tasks to downstream agents without human review at each step.¶

The Model Context Protocol (MCP) provides a common mechanism for connecting agents to external tools, data sources, and APIs. The Agent-to-Agent (A2A) protocol provides complementary support for inter-agent communication and task delegation.¶

These protocols address connectivity and interoperability. They do not, by themselves, define a cryptographic mechanism that binds an agent's authorized action scope to protocol execution, enforces that scope at the protocol boundary, or produces receipts that can be independently verified after the fact.¶

Existing governance approaches often rely on application-layer monitoring, behavioral detection, or post-execution audit logging. In deployments where such governance mechanisms are co-located with the governed agent, a compromised, manipulated, or misconfigured agent may influence the governance outcome.¶

AgentROA addresses this concern by positioning enforcement at the protocol boundary, external to the agent's execution context. The Border Gateway enforcement proxy intercepts MCP tool call requests before they reach the target MCP server, validates the caller's ROA envelope, enforces monotonic scope-narrowing semantics across delegation chains, and produces a cryptographic AER receipt as an intrinsic output of the enforcement decision.¶

The governing question AgentROA answers is: was this agent authorized to perform this specific action under this specific policy envelope at this specific moment — and can that be shown to a third party, including a regulator, auditor, or counter-party, without relying on the agent's own account of execution?¶

1.2. BGP RPKI Analogy

AgentROA draws an analogy from the BGP Route Origin Authorization framework specified in RFC 6480 (RPKI) and RFC 6811 (BGP Prefix Origin Validation).¶

In BGP RPKI, a Route Origin Authorization (ROA) is a cryptographically signed object that states which Autonomous System (AS) is authorized to originate routes for a specific IP address prefix. Border routers validate incoming route announcements against the ROA database before accepting them into the routing table. This prevents route hijacking by ensuring that only authorized ASes can announce specific prefixes.¶

AgentROA applies the same model to AI agent execution:¶

• An Agent ROA envelope states which agent identity is authorized to invoke specific MCP capabilities under specific policy constraints, signed by the authorizing entity.¶

• The Border Gateway validates incoming MCP tool calls against the ROA envelope before forwarding them to the target MCP server. This prevents capability hijacking by ensuring that agents can only invoke capabilities they are explicitly authorized for.¶

• The Agent Route Attestation (ARA) provides per-hop attestation across multi-agent delegation chains, analogous to BGP path validation, ensuring that each hop in a delegation chain operated within its declared scope.¶

The key structural insight from RPKI that AgentROA preserves is the separation of the authorization infrastructure (the ROA envelope and its signing) from the enforcement infrastructure (the Border Gateway). This separation ensures that the enforcement mechanism does not depend on trusting the party being governed.¶

1.3. Scope

This document specifies:¶

• The AgentROA framework architecture and trust boundary model.¶

• The protocol object schemas for ROA envelopes, ARA per-hop attestations, and AER execution receipts.¶

• The monotonic scope-narrowing semantics for multi-agent delegation chains.¶

• The Border Gateway enforcement proxy model for MCP and A2A.¶

Future extensions, including voice-triggered actions and agent identity addressing, are outside the scope of this document.¶

This document does not specify a wire protocol for MCP or A2A, which are governed by their respective specifications. AgentROA operates as an enforcement layer above the MCP OAuth 2.0 authorization mechanism specified in the MCP Authorization specification.¶

3.1. System Components

The AgentROA framework consists of the following components:¶

a) Agent Identity Registry

A registry mapping agent identifiers to cryptographic public keys and capability declarations. Agents register their identity and declared capability scope prior to session initiation.¶

b) Policy Engine

A system that evaluates agent action requests against policy documents and produces ROA envelopes at session initiation. The policy engine is operated by the enterprise or regulated entity deploying the agent.¶

c) Border Gateway

The enforcement proxy described in Section 7. The Border Gateway is the critical enforcement point. It MUST be deployed external to the agent's execution context.¶

e) SCITT Transparency Log (optional)

An append-only transparency log compatible with [I-D.ietf-scitt-architecture] that records AER receipts. When deployed, the SCITT log provides independent third-party verifiability of the agent's execution history.¶

3.2. Trust Boundary Model

The fundamental trust boundary in AgentROA separates two domains:¶

The Agent Execution Domain:

The process or container in which the AI agent framework executes. This domain includes the agent's model inference, tool orchestration, and application-layer code. This domain is considered untrusted from the perspective of the enforcement infrastructure.¶

The Enforcement Domain:

The process or container in which the Border Gateway executes. This domain is separate from and external to the Agent Execution Domain. The Enforcement Domain holds the signing keys for AER receipts and has sole authority to authorize MCP tool calls.¶

This separation is the critical architectural distinction from application-layer governance approaches in which the policy engine and the agents it governs run in the same process. When a governance layer runs in the same process as the agent, a compromised agent may influence the governance outcome.¶

AgentROA assumes that the Border Gateway enforcement proxy runs in a separate process from the governed agent. Deployment MAY use sidecar containers in a Kubernetes pod, separate virtual machines, or hardware-isolated enclaves. The architectural model in this document does not assume a policy decision point that is co-located within the agent's execution context.¶

3.3. Enforcement Proxy Architecture

The Border Gateway implements the following request handling model for MCP tool calls:¶

Step 1: Intercept

The Border Gateway receives the MCP tool call request from the agent. The agent MUST present its current ROA envelope or a valid ARA chain in an authorization header or equivalent protocol metadata field.¶

Step 2: Validate Envelope

The Border Gateway validates: a) The ROA envelope signature (EdDSA verification against the agent's registered public key in the Agent Identity Registry).¶

b) The policy digest (the hash of the policy document in the

envelope MUST match the current policy version for this agent's scope).¶

c) The session identity (the session MUST be active and not

expired).¶

d) The device attestation reference (if required by policy).

Step 3: Validate Scope

The Border Gateway validates that the requested capability (mcp:<server-id>.<tool-name>) is within the authorized scope declared in the ROA envelope or ARA chain.¶

Step 4: Apply No-Loosen Check

If the request originates from a delegation chain (ARA present), the Border Gateway validates the monotonic scope-narrowing invariant: scope(current-ARA) ⊆ scope(parent-ARA-or-ROA). A delegation chain that attempts to expand scope MUST be rejected.¶

Step 5: Produce AER

The Border Gateway produces an Agent Execution Receipt (AER) as an intrinsic output of the enforcement decision. The AER is signed by the Border Gateway's private key. The AER is produced regardless of whether the enforcement outcome is PERMIT or DENY.¶

Step 6: Forward or Reject

If the outcome is PERMIT, the Border Gateway forwards the MCP tool call to the target MCP server. If the outcome is DENY, the Border Gateway returns an authorization error to the agent and records the denial event.¶

Step 7: Write to Ledger

The AER is written to the local audit store and, if configured, to the SCITT transparency log.¶

4.1. ROA Envelope Structure

The AgentROA ROA Envelope is a JSON object with the following top-level fields. All fields are REQUIRED unless noted.¶

schema_version:

String. MUST be "1.0".¶

envelope_id:

String. A unique identifier for this envelope instance, formatted as "env:<random-hex-16>".¶

issued_at:

String. ISO 8601 datetime at which this envelope was issued.¶

expires_at:

String. ISO 8601 datetime at which this envelope expires. Enforcement decisions MUST fail for expired envelopes.¶

session:

Object. Session context. - session_id: String. Unique session identifier. - channel: String. One of: "api", "mcp_client", "voice", "browser", "mobile_app".¶

• agent_id: String. The agent identifier of the governed agent.¶

• device_attestation_ref: String, OPTIONAL. Reference to a device attestation artifact (e.g., WebAuthn assertion).¶

authorized_scope:

Object. The authorized capability scope. - capabilities: Array of Strings. Each element is a capability identifier of the form mcp:<server-id>.<tool-name> or the wildcard form mcp:<server-id>.*.¶

• max_delegation_depth: Integer. Maximum number of delegation hops permitted. MUST be >= 0.¶

• cross_org_permitted: Boolean. Whether delegation across organizational boundaries is permitted.¶

• data_classification_ceiling: String, OPTIONAL. Maximum data classification level accessible.¶

• budget_ceiling: Number, OPTIONAL. Maximum total spend or resource budget permitted for this session, in the unit declared by budget_unit. Downstream delegations MUST NOT declare a budget_ceiling exceeding this value.¶

• budget_unit: String, OPTIONAL. Unit for budget_ceiling (e.g., "USD", "tokens", "api_calls"). Required when budget_ceiling is present.¶

• price_class: Integer, OPTIONAL. Maximum price tier permitted (lower value = lower cost). Downstream delegations MUST NOT declare a price_class exceeding this value.¶

• slo_class: Integer, OPTIONAL. Minimum required service level (higher value = stricter SLO). Downstream delegations MUST NOT declare an slo_class below this value.¶

policy:

Object. Policy binding. - policy_id: String. Identifier of the governing policy. - policy_version: String. Version of the governing policy. - policy_digest: String. SHA-256 hash of the policy document, formatted as "sha256:<hex>".¶

• policy_uri: String, OPTIONAL. URI at which the policy document may be retrieved.¶

authorization:

Object. Authorization requirements. - auth_strength: String. One of: "session_only", "device_bound", "device_bound_with_attestation", "dual_control".¶

• approval_state: String. One of: "pending", "granted", "not_required".¶

• approval_artifact_ref: String, OPTIONAL. Reference to the signed approval artifact.¶

evidence:

Object. Evidence references. - session_hash: String. Hash of the session establishment event.¶

• model_provenance: Array of Strings. Identifiers of the AI models that may execute within this session.¶

signatures:

Array of Objects. One or more signatures. - signer: String. Identifier of the signing entity. - alg: String. Signature algorithm. MUST be "EdDSA". - sig: String. Base64url-encoded signature over the canonical JSON serialization of all fields except "signatures".¶

4.2. Policy Digest Binding

The policy_digest field binds the ROA envelope to a specific version of the governing policy document. This binding serves two purposes:¶

a) Enforcement time validation: The Border Gateway MUST verify

that the policy_digest in the ROA envelope matches the hash of the current policy document for this agent's scope. If the policy has been updated since the envelope was issued, the envelope MUST be rejected and a new envelope issued against the current policy.¶

b) Receipt auditability: The AER receipt includes the policy_digest

from the envelope at the time of the enforcement decision. This allows an auditor reviewing the receipt to retrieve the specific policy version that governed the action, even if the policy has subsequently changed.¶

The canonical policy digest is computed as:¶

policy_digest = "sha256:" || hex(SHA-256(policy_document_bytes))

where policy_document_bytes is the UTF-8 encoding of the policy document in its canonical JSON serialization.¶

4.3. Monotonic Scope-Narrowing Semantics

This section specifies the no-loosen invariant, which is the foundational enforcement property of AgentROA.¶

DEFINITION (Monotonic Scope-Narrowing):

For any two consecutive elements ARA[n-1] and ARA[n] in a delegation chain, the full policy envelope of ARA[n] MUST be no broader than the policy envelope of ARA[n-1] across all constrained dimensions.¶

AgentROA applies a tighten-only partial-order algebra across three distinct constraint dimensions. Each dimension has its own ordering operator, and a violation of ANY dimension MUST cause the Border Gateway to reject the delegation with a DENY AER receipt.¶

Dimension 1 — Capability Scope (set containment):

Let C(x) denote the set of capabilities authorized by envelope or attestation x. Then for all n > 0:¶

      C(ARA[n]) ⊆ C(ARA[n-1])

   where ARA[0] is the root ROA envelope.

An ARA that claims capabilities outside the authorized set of its parent violates this dimension. The denial reason code MUST be SCOPE_EXPANSION_DENIED.¶

Dimension 2 — Budget and Price Class (less-than-or-equal):

Let B(x) denote the budget_ceiling declared in envelope x and P(x) denote the price_class. Then for all n > 0:¶

      B(ARA[n]) ≤ B(ARA[n-1])
      P(ARA[n]) ≤ P(ARA[n-1])

A downstream delegation that raises the budget ceiling or escalates the price class above the parent's declared maximum violates this dimension. The denial reason code MUST be BUDGET_EXPANSION_DENIED.¶

Dimension 3 — Service Level Objective (greater-than-or-equal):

Let S(x) denote the slo_class declared in envelope x, where higher numerical values denote stricter service levels. Then for all n > 0:¶

      S(ARA[n]) ≥ S(ARA[n-1])

A downstream delegation that relaxes the required service level below the parent's declared floor violates this dimension. The denial reason code MUST be SLO_RELAXATION_DENIED.¶

Combined invariant:

The Border Gateway MUST evaluate all three dimensions at each delegation boundary. A delegation envelope that satisfies Dimension 1 but violates Dimension 2 or 3 MUST be rejected. Partial conformance is not sufficient.¶

IMPLEMENTATION REQUIREMENT:

The Border Gateway MUST evaluate this invariant at each delegation boundary. An ARA that violates any dimension of the tighten-only partial-order algebra MUST be rejected. The rejection MUST produce a DENY AER receipt containing the applicable reason code.¶

This invariant prevents privilege escalation through agent delegation chains. An agent that delegates a task to a downstream agent CANNOT grant the downstream agent permissions, budget, or relaxed service levels that the delegating agent does not itself possess.¶

Scope comparison algorithm:¶

The Border Gateway compares capability sets using the following procedure:¶

1. Resolve each capability identifier to its canonical form. Wildcard capabilities (mcp:<server-id>.*) are expanded to include all capabilities advertised by the named MCP server's capability manifest.¶

2. For each capability in the downstream ARA, verify that the capability is present in or subsumed by a wildcard in the parent's capability set.¶

3. If any downstream capability is not subsumed, reject the delegation with a SCOPE_EXPANSION_DENIED error.¶

4. If the downstream ARA declares a budget_ceiling, verify that: budget_ceiling(ARA[n]) ≤ budget_ceiling(ARA[n-1])¶

   If violated, reject with BUDGET_EXPANSION_DENIED.

5. If the downstream ARA declares a price_class, verify that: price_class(ARA[n]) ≤ price_class(ARA[n-1])¶

   If violated, reject with BUDGET_EXPANSION_DENIED.

6. If the downstream ARA declares an slo_class, verify that: slo_class(ARA[n]) ≥ slo_class(ARA[n-1])¶

   If violated, reject with SLO_RELAXATION_DENIED.

4.4. ROA Envelope Signing

ROA envelopes MUST be signed using EdDSA with the Ed25519 curve as specified in [RFC8032]. The signature is computed over the canonical JSON serialization of the envelope excluding the "signatures" field.¶

Canonical JSON serialization follows the JSON Canonicalization Scheme (JCS) as specified in [RFC8785].¶

The signing key MUST correspond to the public key registered in the Agent Identity Registry for the authorizing entity. The key identifier MUST be included in the "signer" field of the signature object.¶

5.1. ARA Object Structure

An Agent Route Attestation (ARA) object is produced at each agent-to-agent delegation boundary. The ARA attests that:¶

a) The delegating agent (the upstream agent) authorized the

downstream agent to act on its behalf.¶

b) The downstream agent's authorized scope is within the

delegating agent's scope (monotonic scope-narrowing).¶

c) The delegation occurred within a valid session.¶

The ARA object has the following structure:¶

schema_version:

String. MUST be "1.0".¶

ara_id:

String. Unique identifier for this ARA, formatted as "ara:<random-hex-16>".¶

issued_at:

String. ISO 8601 datetime at which this ARA was issued.¶

upstream_ref:

Object. Reference to the parent in the delegation chain. - ref_type: String. One of: "roa_envelope", "ara". - ref_id: String. The envelope_id or ara_id of the parent. - ref_digest: String. SHA-256 hash of the parent object's canonical serialization. This creates a hash chain over the delegation path.¶

delegating_agent:

Object. Identity of the delegating agent. - agent_id: String. agent identifier of the delegating agent. - session_id: String. Session identifier of the delegating agent's current session.¶

delegated_agent:

Object. Identity of the agent receiving the delegation. - agent_id: String. agent identifier of the delegated agent. - capability_declaration_ref: String, OPTIONAL. Reference to the delegated agent's declared capability manifest.¶

delegated_scope:

Object. The scope being delegated. MUST satisfy the monotonic scope-narrowing invariant with respect to the parent scope. - capabilities: Array of Strings. Capability identifiers authorized for the delegated agent.¶

• task_context: String, OPTIONAL. Plain-language description of the task being delegated.¶

• max_delegation_depth: Integer. Remaining permitted delegation depth. MUST be strictly less than the parent's max_delegation_depth.¶

policy:

Object. Policy binding for this delegation. - policy_digest: String. MUST match the policy_digest of the root ROA envelope in the delegation chain.¶

• policy_version: String.¶

signatures:

Array of Objects. Signatures as specified in Section 4.4. The ARA MUST be signed by the delegating agent.¶

5.2. Delegation Chain Construction

A delegation chain is constructed as follows:¶

1. The root element is the ROA envelope issued by the Policy Engine at session initiation (ref_type: "roa_envelope").¶

2. Each subsequent element is an ARA object produced at a delegation boundary (ref_type: "ara").¶

3. The chain forms a singly-linked list through the upstream_ref fields. The ref_digest field creates a cryptographic hash chain: each ARA commits to the exact bytes of its parent.¶

4. The chain terminates at the agent currently requesting a MCP tool call.¶

When a Border Gateway receives a MCP tool call, the requesting agent presents its delegation chain (an ordered array of serialized ARA objects, from root to current). The Border Gateway validates the entire chain before making an enforcement decision.¶

Chain length:

The chain length MUST NOT exceed the max_delegation_depth specified in the root ROA envelope.¶

Chain integrity:

The Border Gateway MUST verify: a) The root ROA envelope signature. b) Each ARA signature in the chain. c) The hash chain integrity (ref_digest field). d) The monotonic scope-narrowing invariant at each hop.¶

5.3. Chain Verification Algorithm

The following algorithm MUST be implemented by the Border Gateway:¶

FUNCTION VerifyChain(chain, requested_capability):

  INPUT:
    chain: ordered array [root_envelope, ara_1, ..., ara_n]
    requested_capability: capability identifier being requested

  OUTPUT:
    PERMIT or DENY, with reason

  STEP 1: Verify root envelope
    IF NOT VerifySignature(chain[0]) THEN
      RETURN DENY, "invalid_root_signature"
    IF chain[0].expires_at < NOW() THEN
      RETURN DENY, "envelope_expired"
    current_scope = chain[0].authorized_scope.capabilities
    current_budget = chain[0].authorized_scope.budget_ceiling
    current_slo = chain[0].authorized_scope.slo_class

  STEP 2: Verify hash chain and monotonic narrowing
    FOR i = 1 TO length(chain) - 1:
      ara = chain[i]
      parent = chain[i-1]
      parent_digest = SHA-256(canonical_json(parent))

      IF ara.upstream_ref.ref_digest != parent_digest THEN
        RETURN DENY, "chain_integrity_violation"

      IF NOT VerifySignature(ara) THEN
        RETURN DENY, "invalid_ara_signature_at_hop_" + i

      IF NOT Subset(ara.delegated_scope.capabilities, current_scope) THEN
        RETURN DENY, "scope_expansion_violation_at_hop_" + i

      IF ara.delegated_scope.budget_ceiling IS PRESENT AND
         current_budget IS PRESENT AND
         ara.delegated_scope.budget_ceiling > current_budget THEN
        RETURN DENY, "budget_expansion_denied_at_hop_" + i

      IF ara.delegated_scope.slo_class IS PRESENT AND
         current_slo IS PRESENT AND
         ara.delegated_scope.slo_class < current_slo THEN
        RETURN DENY, "slo_relaxation_denied_at_hop_" + i

      current_scope = ara.delegated_scope.capabilities
      current_budget = ara.delegated_scope.budget_ceiling
      current_slo = ara.delegated_scope.slo_class

• STEP 3: Verify requested capability is in current scope

  IF NOT In(requested_capability, current_scope) THEN RETURN DENY, "capability_not_in_scope"¶

  STEP 4: Verify policy digest consistency

  FOR i = 1 TO length(chain) - 1: IF chain[i].policy.policy_digest != chain[0].policy.policy_digest THEN RETURN DENY, "policy_digest_mismatch_at_hop_" + i¶

RETURN PERMIT¶

END FUNCTION¶

6.1. AER Structure

The Agent Execution Receipt (AER) is produced by the Border Gateway as an intrinsic output of each enforcement decision. The AER is a cryptographically signed artifact intended to provide independently verifiable evidence that a specific enforcement decision was made for a specific agent action under a specific policy at a specific time.¶

The AER is not a log entry. It is a signed commitment produced by the enforcement layer at the moment of the decision. The distinction is architecturally significant: a log entry is produced after the fact by a logging system that may or may not have observed the exact enforcement logic. An AER is produced by the enforcement logic itself, committed before the action executes (for PERMIT) or when the action is rejected (for DENY).¶

AER structure:¶

schema_version:

String. MUST be "1.0".¶

aer_id:

String. Unique receipt identifier, formatted as "aer:<random-hex-16>".¶

produced_at:

String. ISO 8601 datetime at which this receipt was produced. This MUST be the timestamp of the enforcement decision, not the timestamp of execution completion.¶

enforcement_outcome:

String. One of: "permit", "deny".¶

enforcement_mode:

String. One of: "normal", "degraded". "degraded" indicates that the enforcement decision was made against locally cached materials because the Agent Identity Registry or revocation infrastructure was unavailable. Relying parties that require non-degraded receipts SHOULD reject receipts with enforcement_mode "degraded". See Section 9.4.3.¶

denial_reason:

String. Present only when enforcement_outcome is "deny". One of: "invalid_signature", "envelope_expired", "envelope_revoked", "replay_detected", "chain_integrity_violation", "scope_expansion_violation", "budget_expansion_denied", "slo_relaxation_denied", "capability_not_in_scope", "policy_digest_mismatch", "approval_required", "auth_strength_insufficient".¶

session:

Object. Session context. - session_id: String. - agent_id: String. agent identifier of the requesting agent. - device_attestation_ref: String, OPTIONAL.¶

action:

Object. The action subject to enforcement. - capability: String. The capability identifier requested. - mcp_server_id: String. Target MCP server identifier. - mcp_tool_name: String. Target tool name. - input_hash: String. SHA-256 hash of the canonical serialization of the tool call inputs. The inputs themselves are NOT included in the AER.¶

policy:

Object. Policy binding at enforcement time. - policy_id: String. - policy_digest: String. The policy digest from the ROA envelope. This records the exact policy version that governed this enforcement decision.¶

chain_summary:

Object. Summary of the delegation chain. - chain_depth: Integer. Number of hops in the chain. - root_envelope_id: String. The envelope_id of the root ROA envelope.¶

• chain_digest: String. SHA-256 hash of the canonical serialization of the complete delegation chain.¶

border_gateway:

Object. Border Gateway identity. - gateway_id: String. Identifier of the producing Border Gateway instance.¶

• gateway_version: String. Software version.¶

plan_hash:

String, OPTIONAL. A cryptographic hash of the frozen artifacts that governed this enforcement decision, enabling deterministic or tolerance-bounded replay. The plan_hash binds at minimum: the prompt template identifier and version, the model identifier and version, the operator and governance library versions, and the policy-pack identifiers active at enforcement time. When present, it is formatted as "sha256:<hex>" and computed over the canonical JSON serialization of the frozen artifact manifest. Auditors or compliance systems MAY use plan_hash to re-execute the enforcement decision logic against the same frozen inputs and verify that the original allow/deny outcome is reproduced within declared tolerances. Implementations that support deterministic replay SHOULD populate this field.¶

signatures:

Array of Objects. The AER MUST be signed by the Border Gateway's private key using EdDSA/Ed25519. The Border Gateway signing key MUST be registered in the Agent Identity Registry.¶

6.2. Receipt Generation at Enforcement Boundary

The following invariants apply to AER generation:¶

Invariant 1 (Pre-execution commitment):

For enforcement_outcome "permit", the AER MUST be produced and persisted to the local audit store BEFORE the MCP tool call is forwarded to the target MCP server. This ensures that the AER exists regardless of whether the tool call subsequently succeeds or fails.¶

Invariant 2 (Denial receipt):

For enforcement_outcome "deny", the AER MUST be produced and persisted before returning the authorization error to the requesting agent.¶

Invariant 3 (Policy digest capture):

The policy_digest in the AER MUST be captured from the ROA envelope at enforcement time. If the policy has been updated between envelope issuance and enforcement time, the enforcement decision MUST fail with denial reason "policy_digest_mismatch".¶

Invariant 4 (Input hash binding):

The input_hash field MUST be computed from the actual tool call inputs presented at enforcement time. This ensures that the AER receipt is bound to the specific inputs, not merely to the capability identifier.¶

6.3. SCITT Transparency Log Integration

AgentROA AER receipts are designed to be submitted to SCITT- compatible transparency logs [I-D.ietf-scitt-architecture]. When a SCITT log is configured:¶

a) The Border Gateway submits each AER as a SCITT Statement.¶

b) The SCITT log returns a SCITT Receipt (a countersignature

over the log entry).¶

c) The SCITT Receipt is appended to the AER's signatures array.¶

d) The combined AER (with SCITT Receipt) is returned to the

requesting agent and stored in the local audit store.¶

SCITT integration provides three additional properties:¶

1. Third-party verifiability: Any party with the SCITT log's public key can verify the AER without trusting the Border Gateway operator.¶

2. Tamper evidence: The append-only SCITT log structure prevents modification of historical receipts.¶

3. Transparency: Deployments MAY use transparency logs to support independent verification or audit requirements.¶

7.1. Enforcement Proxy Position

The Border Gateway MUST be positioned as follows:

Agent (Execution Domain) | | MCP tool call request | (with ROA envelope or ARA chain in auth header) | v Border Gateway (Enforcement Domain) <-- separate process | | Validated and authorized MCP tool call | (after enforcement decision, AER produced) | v MCP Server¶

The Border Gateway acts as a reverse proxy for MCP connections. From the MCP server's perspective, all requests arrive from the Border Gateway. From the agent's perspective, the Border Gateway is the target endpoint for MCP tool calls.¶

This architecture ensures:

a) The MCP server never receives unauthorized tool calls. b) The enforcement decision is made by infrastructure the agent cannot influence.¶

c) Every tool call — authorized or denied — produces a signed

receipt in the enforcement domain.¶

7.2. MCP Protocol Integration

MCP uses OAuth 2.0 for authorization as specified in the MCP Authorization specification. AgentROA operates as an additional layer above MCP's OAuth authorization mechanism.¶

AgentROA DOES NOT replace MCP's OAuth authorization. Both mechanisms operate in sequence:¶

Step 1 (MCP OAuth): The agent obtains an OAuth access token

scoped to the target MCP server. This token is validated by the MCP server's authorization server.¶

Step 2 (AgentROA envelope): The agent presents its ROA envelope

or ARA chain to the Border Gateway in an authorization header or equivalent protocol metadata field.¶

Step 3 (Border Gateway validation): The Border Gateway performs

the chain verification algorithm specified in Section 5.3 and produces an AER receipt.¶

Step 4 (Forwarding): If the enforcement outcome is PERMIT, the

Border Gateway forwards the tool call to the MCP server with the OAuth token. The AER receipt identifier or receipt object is returned to the agent in response metadata.¶

OAuth token audience validation:

The Border Gateway MUST validate that the OAuth token's audience claim (aud) matches the target MCP server's registered identifier. This prevents token reuse across MCP servers.¶

MCP server registration:

MCP servers that participate in AgentROA enforcement MUST register their capability manifests with the Agent Identity Registry. The capability manifest declares all tools the server exposes, used by the Border Gateway for wildcard scope resolution.¶

7.3. A2A Protocol Integration

When agents communicate using the A2A protocol, AgentROA governs the delegation of tasks between agents.¶

Task delegation over A2A:

When an orchestrating agent delegates a task to a downstream agent over A2A, the orchestrating agent MUST produce an ARA object for the delegation (see Section 5) and include it in the A2A task request.¶

Cross-organization delegation:

When a delegation crosses an organizational boundary, the ROA envelope's cross_org_permitted field MUST be true. Additional cross-organization agreement semantics are outside the scope of this document.¶

7.4. Enforcement Decision Algorithm

The complete enforcement decision algorithm is:¶

FUNCTION Enforce(request):

  INPUT:
    request: MCP tool call request with auth headers

  OUTPUT:
    PERMIT or DENY, with AER receipt

  STEP 1: Extract credentials
    envelope_or_chain = ExtractFromHeader(request)

  STEP 2: Determine chain type
    IF envelope_or_chain is ROA envelope:
      chain = [envelope_or_chain]
    ELSE:
      chain = envelope_or_chain (ARA chain)

  STEP 3: Resolve capability
    capability = mcp:<request.server_id>.<request.tool_name>

  STEP 4: Execute chain verification
    result = VerifyChain(chain, capability)

  STEP 5: Check approval state if required
    IF chain[0].authorization.auth_strength IN
        ["device_bound", "device_bound_with_attestation"]:
      IF chain[0].authorization.approval_state != "granted":
        result = DENY, "approval_required"

  STEP 6: Produce AER
    aer = ProduceAER(
      session = ExtractSession(chain),
      action = ExtractAction(request),
      policy = ExtractPolicy(chain[0]),
      chain_summary = SummarizeChain(chain),
      outcome = result.outcome,
      denial_reason = result.reason
    )
    SignAER(aer, border_gateway_private_key)
    PersistAER(aer)

  STEP 7: Submit to SCITT (if configured)
    IF scitt_enabled:
      scitt_receipt = SubmitToSCITT(aer)
      AppendSignature(aer, scitt_receipt)

  STEP 8: Return
    IF result.outcome == PERMIT:
      ForwardToMCPServer(request)
      RETURN response WITH receipt metadata referencing aer.aer_id
    ELSE:
      RETURN 403 Forbidden WITH receipt metadata referencing
             aer.aer_id

END FUNCTION¶

8.1. Threat Model

AgentROA is designed to protect against the following threats:¶

T1 - Capability escalation through delegation:

A malicious or compromised agent attempts to grant a downstream agent capabilities it does not possess. Mitigation: No-loosen invariant (Section 4.3).¶

T2 - Prompt injection leading to unauthorized tool calls:

An adversary embeds malicious instructions in data the agent processes, causing the agent to attempt tool calls outside its declared scope. Mitigation: Border Gateway scope enforcement (Section 7). The agent cannot invoke capabilities outside its envelope regardless of what instructions it receives.¶

T3 - Replay attacks using captured envelopes:

An attacker captures a valid ROA envelope and replays it to authorize actions not intended by the original issuer. Mitigation: Session identity binding and envelope expiry. The session_id in the envelope MUST be unique per session, and the Border Gateway MUST maintain a replay prevention cache for recently-seen envelope identifiers.¶

T4 - Governance bypass through in-process compromise:

A compromised agent attempts to modify or disable the governance layer. Mitigation: Trust boundary separation (Section 3.2). The Border Gateway runs in a separate process that the agent cannot access.¶

T5 - Policy drift — agent scope expands over time:

An agent's effective capability scope expands through learned skills or updated configuration without the policy envelope being reissued. Mitigation: Policy digest binding (Section 4.2). The policy digest in the ROA envelope is validated against the current policy at each enforcement decision.¶

8.2. Key Management

Private keys used for signing ROA envelopes, ARA objects, and AER receipts MUST be stored in hardware security modules or equivalent secure key storage, or in device-bound key storage for device-bound approval flows.¶

Key rotation:

Implementations SHOULD support periodic rotation of ROA envelope signing keys and Border Gateway signing keys according to local deployment policy and risk requirements.¶

8.3. Implementation Considerations

Implementers SHOULD note that this framework requires the
Border Gateway to process every MCP tool call.  The Border
Gateway is therefore a potential bottleneck and single point
of failure.  Implementations SHOULD provide:

a) High availability deployment (multiple Border Gateway

instances with consistent state).¶

b) Low-latency enforcement appropriate to the deployment, noting

that enforcement occurs at the protocol boundary rather than in application middleware.¶

c) Audit store durability (AER receipts MUST survive Border

Gateway failures).¶

8.4. Revocation and Degraded Operation

This section specifies requirements for ROA envelope revocation and for continued operation when the Border Gateway cannot reach the Agent Identity Registry or policy infrastructure.¶

A.1. ROA Envelope Schema