io.github.mishrasanjeev/grantex MCP Server

<div align="center">

Grantex

Open-Source AI Agent Authorization and Delegated Access

What OAuth 2.0 is to humans, Grantex is to agents.

<br/>

<br/>

Docs | Quickstart | Release JSON | LLM Index | Spec | IETF Draft

<br/>

Ownership: Grantex is owned by Orchestrum Technologies LLP. Inventor and owner: Sanjeev Kumar. Contact: sanjeev@orchestrum.in or mishra.sanjeev@gmail.com.

<br/> <img src="docs/images/flow-diagram.svg" alt="Grantex Protocol Flow" width="100%"/> <br/> </div>

What is Grantex?

Grantex is an open-source delegated authorization protocol and reference implementation for AI agents. It gives each agent a verifiable identity and scoped, time-limited, revocable authority from a human or organization, with multi-agent delegation, service-side verification, and audit records.

Grantex complements OAuth 2.0 and MCP: OAuth handles application and user authorization, MCP connects models to tools, and Grantex proves which agent may perform which action for which principal. Use Grantex when an AI agent acts for a person or organization and a relying service must verify exactly what that agent may do.

For AgenticOrg business onboarding cases, Grantex grants provide delegated tool authority while AgenticOrg owns tenant isolation, case state, the local case-purpose allowlist, human review and the provider call boundary. The last verified AgenticOrg integration does not enforce token-level case purpose or per-case caps. Published Python grantex==0.7.1 contains purpose and cap APIs, but that integration has not been verified to use them. See the governed-case integration boundary and release status. Neither this repository nor an MCP tool grants an agent authority to approve a case as a human.

Agent Prepaid Wallets and x402 v2

Repository source includes principal-controlled prepaid wallets for AI agents. A principal can assign one or multiple wallets, apply assignment, wallet, agent, shared budget-group, principal, and developer controls, constrain recipients, resource origins, actions, merchants, purposes, projects, and cost centers, require an exact human approval, govern reload velocity, and block one assignment, one wallet, or all wallets available to an agent. Authorizations reserve value atomically in PostgreSQL and are bound to the agent's DPoP OAuth identity and the policy-evaluated semantic context.

@grantex/x402 uses official x402 v2 PAYMENT-REQUIRED, PAYMENT-SIGNATURE, and PAYMENT-RESPONSE messages. The old simulated X-Payment-Proof path is not used. sandbox_ledger is implemented end to end; external custody records deliberately fail closed unless a configured provider can verify funding and settlement. The source checkout now includes opt-in Base native USDC EIP-3009 payments with policy-gated signing, verified funding, durable signature retries and finalized-chain reconciliation. Python 0.7.1 and Go v0.4.2 publish the EVM response and reconciliation APIs; neither adds an automatic HTTP payment wrapper. Published TypeScript 0.8.1 and x402 0.4.1 include the opt-in automatic Base 402/sign/retry flow. All four releases are registry verified. See the Base USDC custody guide, including the limit that blocking cannot recall an already issued on-chain signature.

import { PrepaidWalletAgentClient } from '@grantex/sdk';
import { createX402Agent } from '@grantex/x402';

const walletAgent = new PrepaidWalletAgentClient({ oauthClient, accessToken });
const x402 = createX402Agent({
  authorizePayment: walletAgent.x402Authorizer,
  // walletId is optional; omit it for policy-based wallet selection.
});

const logicalPaymentId = 'order_01JZ8Y6Q2M4N7P9T';
const response = await x402.fetch('https://merchant.example/paid-resource', {
  // The merchant uses this header to recover a response lost after settlement.
  headers: { 'Idempotency-Key': logicalPaymentId },
  // Grantex uses this option to recover a reservation response lost before settlement.
  idempotencyKey: logicalPaymentId,
});

For side-effecting resources, the merchant must durably cache the business result under Idempotency-Key. Grantex settlement is idempotent, but a settled payment authorization cannot be verified or execute protected work a second time.

The OAuth grant must include wallet:spend and the exact merchant action scope advertised as extra.grantexScope; wallet assignment policy can restrict that human-approved authority but cannot add to it.

Where Grantex fits

Grantex owns delegated agent identity, semantic spend policy, cross-wallet budgets, exact approvals, atomic reservations, stop controls, and audit. An AgenticOrg deployment or another agent runtime owns orchestration and the human interaction. The wallet issuer or custodian remains responsible for KYC/KYB, AML and sanctions controls, custody, network authorization, MCC/geography/card controls, settlement, FX, refunds, disputes, fraud, and reconciliation. The merchant remains responsible for server-trusted price/payee data, order idempotency, settlement gating, and delivery of the paid result.

See Agent Wallet Governance for the complete responsibility matrix, policy composition, exact approval protocol, and honest residual gap list.

The managed clients are implemented in registry-verified @grantex/sdk@0.8.1, @grantex/x402@0.4.1, Python grantex==0.7.1, and Go v0.4.2. External custody, principal notification delivery, and merchant result idempotency remain operator responsibilities. See the x402 integration guide, wallet lifecycle, and production-readiness guide.

Self-hosted prepaid-wallet dependencies

Running the auth-service does not provision every dependency needed for a real-money product. Self-hosting operators must explicitly provide and test:

• PostgreSQL migrations 091_agent_prepaid_wallets.sql and 092_layered_wallet_spend_controls.sql, backups, restore, and reconciliation for the append-only wallet ledger and policy decisions;
• HTTPS ingress for the exact OAuth audience, including /v1/prepaid-wallets and /v1/prepaid-wallets/** (a static-host 404 breaks agent wallet access even while the origin service is healthy);
• a provider-specific custody, funding, settlement, webhook-deduplication, and recovery adapter before using external wallets; without one, Grantex deliberately returns 503 CUSTODY_ADAPTER_UNAVAILABLE;
• an SSE/WebSocket notification bridge for reload events when a principal must be notified by email, SMS, Slack, WhatsApp, or another external channel; no built-in wallet-email delivery is claimed;
• a merchant-side durable result cache keyed by the HTTP Idempotency-Key for side effects after settlement;
• independent security, provider, reconciliation, incident-response, and applicable legal/regulatory review.

The server deployment and each SDK release are separate. Registry consumers should verify an exact version before installing it rather than inferring availability from repository manifests. The production-readiness guide contains the full hosting checklist and a maintainer-only PowerShell publication runbook.

Supply-chain and license controls

The security workflow audits every tracked npm lockfile, every tracked Python project and requirements surface, and all three Go modules. Dependabot coverage is checked against the repository manifests, external GitHub Actions and container images are pinned to immutable digests, and unreviewed npm license identifiers fail CI. Run npm run audit:supply-chain and npm run audit:python locally before a release.

Apache-2.0 covers Grantex source, not third-party dependencies. See THIRD_PARTY_NOTICES.md for the current Sharp/libvips LGPL and caniuse-lite CC-BY distribution notes, and the supply-chain security guide for the operator checklist. This inventory is an engineering control, not a legal non-infringement opinion; distributors must review the exact artifacts they ship.

OAuth Agent Grants Profile

The hosted auth service and repository TypeScript client implement the three roles in candidate draft-mishra-oauth-agent-grants-03. The tested profile requires PAR, PKCE S256, DPoP sender constraints, RFC 9207 response issuer validation, five-minute access tokens, rotating refresh tokens with family replay revocation, 300-second exact-request recovery when a refresh response is lost, AES-256-GCM encrypted recovery state with expiry cleanup, same-resource RFC 8693 attenuation, and RFC 7009 revocation.

Discovery and endpoints	URL
Authorization-server metadata	https://grantex.dev/.well-known/oauth-authorization-server
PAR / authorize / token / revoke	https://grantex.dev/oauth/{par,authorize,token,revoke}
Implementation guide	docs.grantex.dev/guides/oauth-agent-grants
Evidence	Implementation report and 30 behavioral vectors

Verification completed with 2,138 auth-service tests, 456 TypeScript SDK tests, and 255 tests across 23 sequential Docker E2E files. This is self-assessed evidence for the tested Grantex configuration, not independent interoperability certification, OAuth Working Group adoption, or IETF endorsement. The updated OAuthAgentClient is published in registry-verified @grantex/sdk@0.8.1.

Revision -02 remains the current Datatracker publication. Candidate -03 must not be uploaded before 2026-09-09 and requires a fresh explicit approval after the scheduled final review.

Open Agentic Commerce Protocol (OACP) Authority

Open Agentic Commerce Protocol (OACP) is Grantex's agentic-commerce trust and artifact-authority layer. Grantex governs OACP policy, internal artifact issuance or refusal, verification, and compatibility adapters. AgenticOrg owns buyer and seller AI-agent runtime, merchant self-service onboarding, Shopify connector runtime, future merchant connector setup intent, buyer sessions, channel bridges, OACP cache, and provider-owned capability verification.

Merchant systems such as Shopify, future WooCommerce/ERP sources, POS systems, and provider systems remain the source of record. Provider, bank, POS, and payment rails own mandate, payment, and in-store execution. Grantex signs and verifies artifacts; it is not a merchant connector runtime or a toll booth for every buyer and seller message.

flowchart LR
  merchant[Shopify, future ERP/WooCommerce, POS, provider systems] --> agentic[AgenticOrg buyer and seller runtime]
  agentic -->|redacted authority request| grantex[Grantex OACP authority]
  grantex -->|OACP artifacts or blockers| agentic
  agentic --> buyer[Buyer surfaces]
  agentic -->|capability check or handoff| provider[Pine Labs Plural/P3P, bank/POS/provider rails]

Area	Current posture
Grantex C6Z authority route	Implemented at POST /v1/commerce/oacp/c6z/authority-requests for allowlisted AgenticOrg tenants.
Artifact families	11 internal OACP families are issued or refused with source lineage, TTL, freshness, revocation posture, blocked capabilities, non-sensitive evidence refs, and signature metadata.
Protocol adapters	Schema.org, UCP-style, ACP-style, AP2-style, A2A, MCP, and OpenAPI mappings are compatibility mappings derived from OACP artifacts.
AgenticOrg runtime	Merchant self-service config, Seller onboarding, Shopify sync, future connector/provider intent capture, cache, buyer Q&A, bridges, and provider capability verification live in AgenticOrg.
Payment/order/POS execution	Outside OACP artifact authority. Provider, POS, and merchant systems must execute and confirm; agents must not invent success.
Historical Commerce V1 docs	Retained for context, but superseded for the AgenticOrg OACP runtime split.

Start with the OACP runtime launch closure PRD, OACP authority overview, merchant self-service config boundary, truth inventory, AgenticOrg integration guide, POS bridge boundary, and operator runbook. The older Commerce V1 overview remains historical/contextual and should not be used to imply that Grantex owns AgenticOrg merchant connector runtime.

Agent Trust Registry: accredited issuers and Agent Passports

The registry lets a relying party, such as a merchant, a payment service provider or an API, check an AI agent before it acts. It answers three questions: which software the agent is, which key it holds, and which accredited issuer vouches for it. The answer comes as a computed trust level, not a claim the agent makes about itself. The grant still says what the agent may do; the passport says who the agent is. See Passport vs grant.

Piece	What is on main
Accredited issuers	The registry operator accredits issuers, each with a static JWKS, a status_list_base and the trust-mark types it may attest. They are managed with POST and PATCH /v1/registry/issuers under an operator key (REGISTRY_OPERATOR_API_KEYS). An issuer can be suspended; while it is, its attestations stop counting.
Agent keys	Each agent has a key history (/v1/agents/{id}/keys). A key must prove possession by answering a challenge before anything can attest it. Keys can be rotated with an overlap, or reported compromised, which is final.
Attestations	An accredited issuer posts a compact JWS to POST /v1/registry/attestations. The registry checks the issuer, the signature, the agent and its proven key, the hash rule, the validity times and the issuer's own Token Status List. Every refusal carries a stable code and a reason.
Trust levels	basic, verified, attested or attested_verified, with flags such as key_compromised, issuer_suspended and attestation_expiring. Anything suspended reads basic: the registry fails closed.
Acceptance status lists	The registry publishes its own acceptance of each attestation in two formats: as a Token Status List (/status/attestations/{list}) and as a Bitstring Status List.
Lookup and manifest	GET /v1/registry/agents/{did} and the thumbprint and credential forms return a minimised record. /.well-known/agent-registry.json is a signed manifest of the issuers, their keys, the trust-mark taxonomy and the status lists, for relying parties without Federation support.
Passport-bound grants	POST /v1/authorize can take an Agent Passport (SD-JWT VC). The grant is bound to the passport's key and to the registry's acceptance entry, and the binding is checked again at every code exchange and refresh. A passport-bound grant cannot be delegated yet.

New behaviour on existing paths is off by default. Turn it on per deployment:

• REGISTRY_PUBLIC_ENDPOINTS_ENABLED serves the unauthenticated reads: the issuer list, the status lists, the lookup without an API key and the manifest.
• PASSPORT_BOUND_GRANTS_ENABLED turns on passport-bound grants.
• AGENT_KEY_HISTORY_MIRROR_ENABLED makes POST and PATCH /v1/agents write the agent's key into the key history, and refuse a key that is compromised or belongs to another agent.
• REGISTRY_ATTESTATION_EDDSA_ENABLED accepts EdDSA attestations as well as ES256.

Self-hosting lists every variable. Guides for each audience:

• Becoming an accredited issuer
• Registering agents, for providers
• Verifying agents, for relying parties

The specifications are registry federation, attestations, agent keys, Agent Passport 1.0, passport binding and verification. This Agent Passport, an SD-JWT VC issued by an accredited issuer, is separate from the MPP AgentPassportCredential described below.

Current Releases

Grantex components are independently versioned. The protocol specification remains v1.0 Final; SDK, MCP package, and roadmap milestone versions are separate release lines and do not represent a monorepo-wide version.

Current public releases and repository versions, verified 2026-09-30:

The enforcement releases are TypeScript 0.8.1, Python 0.7.1, CLI 0.4.2, gateway/adapters/Strands 0.2.1 and MCP Auth 4.0.0. See the complete migration guide before upgrading: Node.js requirements, audience binding, capped-call amounts and online revocation defaults are breaking changes.

TypeScript 0.8.1 and Python 0.7.1 include service-side enforce() audience binding, denial of capped calls without trusted amounts, and online revocation by default. Register the agent's resource servers before audience-bound authorization, configure the intended audience at the service boundary, and supply a validated amount for every capped call. Offline verification alone does not prove current revocation; integrations must explicitly wire their current-state enforcement. Node.js 22.12+ is required for the changed npm packages. See release limitations.

Component	Published version	Repository version	Reproducible install
TypeScript SDK	@grantex/sdk 0.8.1	0.8.1	npm install @grantex/sdk@0.8.1
x402 Payment Protocol	@grantex/x402 0.4.1	0.4.1	npm install @grantex/x402@0.4.1 @grantex/sdk@0.8.1
Python SDK	grantex 0.7.1	-	python -m pip install grantex==0.7.1
Go SDK	github.com/mishrasanjeev/grantex-go v0.4.2 (Go 1.26.1+)	-	go get github.com/mishrasanjeev/grantex-go@v0.4.2
MCP Authorization Server	@grantex/mcp-auth 4.0.0	4.0.0	npm install @grantex/mcp-auth@4.0.0 @grantex/sdk@0.8.1
CLI	@grantex/cli 0.4.2	0.4.2	npm install -g @grantex/cli@0.4.2
Gateway	@grantex/gateway 0.2.1	0.2.1	npm install @grantex/gateway@0.2.1 @grantex/sdk@0.8.1
Service adapters	@grantex/adapters 0.2.1	0.2.1	npm install @grantex/adapters@0.2.1 @grantex/sdk@0.8.1
Strands TypeScript	@grantex/strands 0.2.1	0.2.1	npm install @grantex/strands@0.2.1 @grantex/sdk@0.8.1
Strands Python	grantex-strands 0.2.1	0.2.1	python -m pip install grantex-strands==0.2.1 grantex==0.7.1

Python registry hashes, public-file tests and normal pinned PyPI index installations are verified on the workstation and in Docker. See the validation report.

Deployment responsibilities: MCP Auth 4.0.0 supports durable shared storage, rendered consent and corrected code handoff. Configure shared state, resource audience, a required host-authenticated human-principal resolver and trusted current-grant verification before production use. A consent page does not itself authenticate a human or provision passkeys. See the release-status guide.

Repository development status: the auth service enforces Redis-backed Free/Pro/Enterprise developer budgets of 100/500/2,000 requests per minute on API-key routes handled by the standard auth plugin; revoking and the emergency stop have a budget of their own, as do the revocation feed and status reads. Custom-auth quota policy remains open and the managed-service rollout remains independent of SDK publication.

Omit a version pin to install the registry's current latest release. See the release-status documentation, COMPATIBILITY.md for the full package matrix, and CHANGELOG.md for release notes.

• @grantex/gemma: Offline consent bundles and on-device verification examples
• MCP Authorization Server (@grantex/mcp-auth): OAuth 2.1 + PKCE endpoints with durable storage adapters, rendered consent, authenticated human-principal resolution, resource binding and current-grant verification; version 4 requires host authentication configuration
• MCP Tool Server (@grantex/mcp): Agent-facing Grantex tools for MCP clients
• @grantex/dpdp: DPDP Act 2023 and EU AI Act control mappings
• Trust Registry: Public DID verification registry — grantex.dev/registry
• grantex verify: Token inspection CLI — no account needed
• Agent CLI Skills: One-command SKILL.md installation for Hermes, OpenClaw, and portable Agent Skills clients
• Anomaly Detection: Four implemented SQL-backed checks, lifecycle APIs, and stored rule/channel configuration; notification delivery requires a host worker

---

SDK quickstart

Execution authority and human identity

Signature verification alone does not establish current permission or prove that the executing agent belongs to the authenticated human. Published TypeScript SDK 0.8.1 adds optional per-invocation issuer checks and trusted principal/agent bindings. Python, Go and execution-integration changes remain source candidates. Existing offline defaults are unchanged. Human authentication and live consent remain issuer/host responsibilities. Action-bound decisions and cumulative spend controls must be enforced using the actual execution inputs, not inferred from a valid JWT. See SDK execution authority for setup, package boundaries and release requirements.

npm install @grantex/sdk@0.8.1

import { Grantex, verifyGrantToken } from '@grantex/sdk';
const gx = new Grantex({ apiKey: process.env.GRANTEX_API_KEY });

// 1. Register an agent, then request authorization from a user
const agent = await gx.agents.register({
  name: 'quickstart-agent',
  description: 'Grantex quickstart agent',
  scopes: ['calendar:read', 'email:send'],
});
const auth = await gx.authorize({
  agentId: agent.id,
  userId: 'user-456',
  scopes: ['calendar:read', 'email:send'],
});

// Live mode requires consent at this URL and returns the code to your callback.
// Sandbox or policy auto-approval can return the code immediately.
if (!auth.code) {
  console.log(`Approve access at: ${auth.consentUrl}`);
} else {
  // 2. Exchange the authorization code for a scoped, signed JWT
  const { grantToken } = await gx.tokens.exchange({ code: auth.code, agentId: agent.id });

  // 3. Verify locally using the issuer's published JWKS
  const grant = await verifyGrantToken(grantToken, {
    jwksUri: 'https://api.grantex.dev/.well-known/jwks.json',
  });
  console.log(grant.scopes); // ['calendar:read', 'email:send']
}

python -m pip install grantex==0.7.1               # Python SDK
go get github.com/mishrasanjeev/grantex-go@v0.4.2 # Go SDK (Go 1.26.1+)
npm install @grantex/mcp-auth@4.0.0 @grantex/sdk@0.8.1 # MCP endpoint evaluation
npm install -g @grantex/cli@0.4.2                   # Optional CLI tooling

Hermes, OpenClaw, and any agent CLI

Shell-capable agents use the same JSON-first CLI; no agent-specific SDK is required:

npm install -g @grantex/cli@0.4.2
grantex agent install --target openclaw  # writes ./skills
grantex agent install --target hermes    # writes ~/.hermes/skills/grantex
grantex agent install --target portable  # writes ./.agents/skills

The bundle installs use-grantex-cli for delegated-authorization operations and integrate-grantex for service-boundary implementation work. For a different host, use grantex agent install --dir /path/to/skills. Prefer --env, --file, or --stdin token inputs and keep final enforcement inside the protected service.

35 packages across TypeScript, Python, and Go. Integrations for Anthropic SDK, LangChain, OpenAI Agents SDK, Google ADK, Strands Agents SDK, CrewAI, Vercel AI, AutoGen, MCP, Express.js, FastAPI, and Terraform. Use the compatibility matrix for versions, the changelog for release notes, and GitHub Actions for current CI status. Fully self-hostable. Apache 2.0. The new relying-party Python verifier remains private source, not a registry release.

---

The Problem

AI agents are booking travel, sending emails, deploying code, and spending money — on behalf of real humans. But:

• No scoping — agents get the same access as the key owner
• No consent — users never approve what the agent can do
• No per-agent identity — you know the key was used, but not which agent or why
• No revocation granularity — one agent misbehaves, rotate the key, kill everything
• No delegation control — Agent A calls Agent B? Copy-paste credentials
• No spending limits — an agent with a cloud API key can provision unlimited resources

OAuth and IAM provide essential foundations, but many agent deployments still rely on shared credentials that do not identify the individual agent or encode its delegated authority.

---

How It Works

<img src="docs/images/flow-diagram.svg" alt="Grantex Protocol Flow" width="100%"/>

---

Quickstart

1. Register your agent

import { Grantex } from '@grantex/sdk';

const grantex = new Grantex({ apiKey: process.env.GRANTEX_API_KEY });

const agent = await grantex.agents.register({
  name: 'travel-booker',
  description: 'Books flights and hotels on behalf of users',
  scopes: ['calendar:read', 'payments:initiate:max_500', 'email:send'],
});

console.log(agent.did);
// → did:grantex:ag_01HXYZ123abc...

2. Request authorization from a user

const authRequest = await grantex.authorize({
  agentId: agent.id,
  userId: 'user_abc123',       // your app's user identifier
  scopes: ['calendar:read', 'payments:initiate:max_500'],
  expiresIn: '24h',
  redirectUri: 'https://yourapp.com/auth/callback',
});

// Redirect user to authRequest.consentUrl
// Grantex handles the consent UI — plain language, mobile-first
console.log(authRequest.consentUrl);
// → https://consent.grantex.dev/authorize?req=eyJ...

3. Exchange the authorization code for a grant token

// After user approves, your redirectUri receives a `code`.
// Exchange it for a signed grant token (RS256 JWT):
const token = await grantex.tokens.exchange({
  code,                  // from the redirect callback
  agentId: agent.id,
});

console.log(token.grantToken);  // RS256 JWT — pass this to your agent
console.log(token.scopes);      // ['calendar:read', 'payments:initiate:max_500']
console.log(token.grantId);     // 'grnt_01HXYZ...'
console.log(token.refreshToken); // store securely for active-grant rotation

Refresh tokens are single-use and rotate on every accepted refresh. If the refresh HTTP response is lost after the server commits, retry the same previous refresh token immediately; Grantex can recover the already-rotated token pair for five minutes (300 seconds). Refresh does not extend token.expiresAt; after the grant expires, start a new authorization request.

4. Verify the token and use it

// Verify locally after retrieving the issuer's published JWKS
import { verifyGrantToken } from '@grantex/sdk';

const grant = await verifyGrantToken(token.grantToken, {
  jwksUri: 'https://api.grantex.dev/.well-known/jwks.json',
  requiredScopes: ['calendar:read'],
});
console.log(grant.principalId); // 'user_abc123'
console.log(grant.scopes);     // ['calendar:read', 'payments:initiate:max_500']

// Pass to your agent — it's now authorized
await travelAgent.run({ grantToken: token.grantToken, task: 'Book cheapest flight to Delhi on March 1' });

5. Record the action at the execution boundary

// At the trusted execution boundary: explicitly record the outcome
await grantex.audit.log({
  agentId: agent.id,
  agentDid: agent.did,
  grantId: token.grantId,
  principalId: authRequest.principalId,
  action: 'payment.initiated',
  status: 'success',
  metadata: { amount: 420, currency: 'USD', merchant: 'Air India' },
});

6. Verify a token (service-side)

// In any service that receives agent requests — no Grantex account needed
import { verifyGrantToken } from '@grantex/sdk';

const grant = await verifyGrantToken(token.grantToken, {
  jwksUri: 'https://api.grantex.dev/.well-known/jwks.json',
  requiredScopes: ['payments:initiate'],
});
// Throws if the token is expired, tampered with, has invalid claims, or lacks required scopes.
// Use grantex.tokens.verify(token.grantToken) when you also need current revocation status.

7. Give users control over their permissions

// Generate a short-lived link for the end-user to view & revoke agent access
const session = await grantex.principalSessions.create({
  principalId: 'user_abc123',
  expiresIn: '2h',
});
// Send session.dashboardUrl to the user via email, in-app notification, etc.
// The short-lived session token is carried in the URL fragment, not the query string.

---

Python SDK

import os

from grantex import Grantex, AuthorizeParams, ExchangeTokenParams

client = Grantex(api_key=os.environ["GRANTEX_API_KEY"])

# Register agent
agent = client.agents.register(
    name="finance-agent",
    scopes=["transactions:read", "payments:initiate:max_100"],
)

# Authorize a user
auth = client.authorize(AuthorizeParams(
    agent_id=agent.id,
    user_id="user_abc123",
    scopes=["transactions:read", "payments:initiate:max_100"],
))
# Redirect user to auth.consent_url — they approve in plain language

# Exchange the authorization code for a grant token
token = client.tokens.exchange(ExchangeTokenParams(code=code, agent_id=agent.id))

# Verify locally after retrieving the issuer's published JWKS
from grantex import verify_grant_token, VerifyGrantTokenOptions

grant = verify_grant_token(token.grant_token, VerifyGrantTokenOptions(
    jwks_uri="https://api.grantex.dev/.well-known/jwks.json",
))
print(grant.scopes)  # ('transactions:read', 'payments:initiate:max_100')

# Log an action
client.audit.log(
    agent_id=agent.id,
    agent_did=agent.did,
    grant_id=token.grant_id,
    principal_id=auth.principal_id,
    action="transaction.read",
    status="success",
    metadata={"account_last4": "4242"},
)

---

The Grant Token

Grantex tokens are standard JWTs (RS256) extended with agent-specific claims. Any service can verify their signatures locally using the issuer's published JWKS. The provided verifiers retrieve those keys from the configured JWKS URL, so applications should account for network availability, caching, and key rotation:

{
  "iss": "https://grantex.dev",
  "sub": "user_abc123",
  "agt": "did:grantex:ag_01HXYZ123abc",
  "dev": "org_yourcompany",
  "scp": ["calendar:read", "payments:initiate:max_500"],
  "iat": 1709000000,
  "exp": 1709086400,
  "jti": "tok_01HXYZ987xyz",
  "grnt": "grnt_01HXYZ456def"
}

Claim	Meaning
sub	The end-user who authorized this agent
agt	The agent's DID — cryptographically verifiable identity
dev	The developer org that built the agent
scp	Exact scopes granted — services should check these
jti	Unique token ID — used for grant-state and revocation checks
grnt	Grant record ID — links token to the persisted grant
aud	Intended audience (optional) — services should reject tokens with a mismatched aud

Delegation claims (present on sub-agent tokens):

Claim	Meaning
parentAgt	DID of the parent agent that spawned this sub-agent
parentGrnt	Grant ID of the parent grant — full delegation chain is traceable
delegationDepth	How many hops from the root grant (root = 0)

---

Multi-Agent Delegation

Grantex supports multi-agent pipelines where a root agent spawns sub-agents with narrower scopes. Sub-agent tokens carry a full delegation chain that any service can inspect.

// Root agent has a grant for ['calendar:read', 'calendar:write', 'email:send']
// It spawns a sub-agent that only needs calendar read access

const delegated = await grantex.grants.delegate({
  parentGrantToken: rootGrantToken,   // root agent's token
  subAgentId: subAgent.id,            // sub-agent to authorize
  scopes: ['calendar:read'],          // must be ⊆ parent scopes
  expiresIn: '1h',                    // capped at parent token's expiry
});

// delegated.grantToken is a fully signed JWT with:
//   parentAgt, parentGrnt, delegationDepth = 1

# Python equivalent
delegated = grantex.grants.delegate(
    parent_grant_token=root_grant_token,
    sub_agent_id=sub_agent.id,
    scopes=["calendar:read"],
    expires_in="1h",
)

Constraints enforced by the protocol:

• Sub-agent scopes must be a strict subset of the parent's scopes — scope escalation is rejected with 400
• Sub-agent token expiry is min(parent expiry, requested expiry) — sub-agents can never outlive their parent
• Revoking a root grant cascades to all descendant grants atomically

---

Advanced Features

<details> <summary><strong>Enterprise SSO</strong> - OIDC and SAML 2.0, plus an LDAP direct-bind preview</summary>

Enterprise SSO

Grantex provides OIDC and SAML 2.0 enterprise SSO with multiple identity-provider connections, email-domain routing, JIT provisioning, and group-to-scope mapping from identity-provider claims. The LDAP surface is a direct-bind preview: it authenticates a supplied directory identity but does not search directories or retrieve LDAP groups. The hosted dashboard sign-in currently supports OIDC; SAML and LDAP callback flows are for custom integrations.

Key capabilities:

• Multi-IdP connections - Configure multiple OIDC and SAML 2.0 identity providers per organization; LDAP connection records support the direct-bind preview
• OIDC Discovery + JWKS verification — Automatic endpoint discovery and cryptographic ID token verification
• SAML 2.0 — Full SAML response parsing with certificate-based signature verification
• LDAP / Active Directory preview - Direct-bind authentication only; directory search and LDAP group retrieval are not implemented
• Domain-based routing — Automatically route users to the correct IdP based on their email domain
• JIT provisioning — Auto-create or update principals on first SSO login
• Group-to-scope mapping - Map OIDC or SAML group/role claims to Grantex scopes; this does not retrieve LDAP groups
• Human SSO enforcement — With the server feature enabled and an organization opt-in, require a matching SSO session for hosted dashboard sign-in and principal consent; live consent still requires a passkey. Machine API keys remain valid for API automation.
• Session management — Track, list, and revoke active SSO sessions

TypeScript

// Create an OIDC connection
const conn = await grantex.sso.createConnection({
  name: 'Okta Production',
  protocol: 'oidc',
  issuerUrl: 'https://mycompany.okta.com',
  clientId: 'your-client-id',
  clientSecret: 'your-client-secret',
  domains: ['mycompany.com'],
  jitProvisioning: true,
  groupAttribute: 'groups',
  groupMappings: { Engineering: ['read', 'write', 'deploy'], Admins: ['admin'] },
  defaultScopes: ['read'],
});

// Create a SAML 2.0 connection
await grantex.sso.createConnection({
  name: 'Azure AD SAML',
  protocol: 'saml',
  idpEntityId: 'https://sts.windows.net/tenant-id/',
  idpSsoUrl: 'https://login.microsoftonline.com/tenant-id/saml2',
  idpCertificate: '-----BEGIN CERTIFICATE-----\n...\n-----END CERTIFICATE-----',
  spEntityId: 'urn:grantex:mycompany',
  spAcsUrl: 'https://myapp.com/sso/callback/saml',
  domains: ['mycompany.com'],
});

// Enforce SSO for the organization
// First complete an OIDC login that maps an administrator to the admin scope.
// The server must have SSO_HUMAN_ENFORCEMENT_ENABLED=true.
await grantex.sso.setEnforcement({ enforce: true });

// Handle OIDC callback with verified ID token
const result = await grantex.sso.handleOidcCallback({ code, state });
console.log(result.email, result.mappedScopes, result.sessionId);
// result.sessionToken is an opaque bearer credential when enabled. Keep it secret.

// List and revoke sessions
const { sessions } = await grantex.sso.listSessions();
await grantex.sso.revokeSession(sessions[0].id);

Python

# Create an OIDC connection
conn = client.sso.create_connection(CreateSsoConnectionParams(
    name="Okta Production",
    protocol="oidc",
    issuer_url="https://mycompany.okta.com",
    client_id="your-client-id",
    client_secret="your-client-secret",
    domains=["mycompany.com"],
    jit_provisioning=True,
    group_attribute="groups",
    group_mappings={"Engineering": ["read", "write", "deploy"]},
))

# Handle OIDC callback
result = client.sso.handle_oidc_callback(SsoOidcCallbackParams(code=code, state=state))
print(result.email, result.mapped_scopes, result.session_id)

CLI

# Create a connection
grantex sso connections create --name "Okta" --protocol oidc \
  --issuer-url https://mycompany.okta.com \
  --client-id $CLIENT_ID --client-secret $CLIENT_SECRET \
  --domains mycompany.com --jit-provisioning

# List connections
grantex sso connections list

# Test connectivity
grantex sso connections test sso_01HXYZ...

# Enforce SSO
grantex sso enforce --enable

Enabling invalidates prior browser SSO sessions and safely migrates only unambiguous, session-proven JIT identities. If a principal was mapped to multiple IdP subjects, enablement fails with SSO_IDENTITY_CONFLICT until resolved. An active JIT-enabled OIDC connection and a recent successful admin SSO login are required. The developer API key is a deliberate machine-access exception: it can still call management APIs and disable enforcement, so store it as a privileged secret. See the enterprise SSO guide for the rollout and local signed-IdP test procedure.

</details> <details> <summary><strong>FIDO2 / WebAuthn</strong> — passkey-based consent verification</summary>

FIDO2 / WebAuthn

Grantex supports passkey-based human presence verification using FIDO2/WebAuthn. Live-mode consent requires a passkey assertion; sandbox-mode accounts can opt in with fidoRequired: true. The assertion confirms use of a registered credential during the consent flow. A passkey-required sandbox request stays pending just like live consent: neither /v1/authorize nor OAuth PAR auto-approves it, and developer-key shortcuts cannot approve or deny it.

How It Works

1. Developer configures FIDO — Live mode already requires it; set fidoRequired: true via PATCH /v1/me to require it in sandbox mode too.
2. Application enrolls the customer — After authenticating the customer, the application issues a one-use hosted enrollment link for their exact principal ID. The consent URL alone cannot register a passkey.
3. User authenticates on consent — On subsequent authorization requests, the user completes a WebAuthn assertion challenge instead of a simple button click
4. Consent requires verified presence — With portable evidence enabled, Grantex stores the verified assertion on the pending request, binds it to the grant, and puts a signed digest reference in the grant token. An opt-in VC carries the assertion verification inputs.

SDK Usage

Hosted enrollment is included in published @grantex/sdk@0.8.1, grantex==0.7.1, and github.com/mishrasanjeev/grantex-go@v0.4.2. The Grantex-hosted service has PASSKEY_ENROLLMENT_ENABLED=true and uses grantex.dev as its WebAuthn RP ID; self-hosted operators must configure their own HTTPS origin and enable the flag. Publishing an SDK does not enable it on a server.

// Enable FIDO for your developer account
await grantex.updateSettings({ fidoRequired: true, fidoRpName: 'My App' });

// Server-side, after authenticating the end customer:
const { enrollmentUrl } = await grantex.webauthn.createEnrollmentSession({
  principalId: authenticatedCustomer.grantexPrincipalId,
});
// Show the one-use link only to that customer. The hosted page handles WebAuthn.

// List and manage credentials
const { credentials: creds } = await grantex.webauthn.listCredentials('user_abc123');
await grantex.webauthn.deleteCredential(credentialId);

# Enable FIDO for your developer account
from grantex import UpdateDeveloperSettingsParams

client.update_settings(UpdateDeveloperSettingsParams(
    fido_required=True,
    fido_rp_name="My App",
))

# After your app authenticates the customer, issue a one-use hosted link.
session = client.webauthn.create_enrollment_session(principal_id="user_abc123")
enrollment_url = session.enrollment_url

# List and manage credentials
creds = client.webauthn.list_credentials("user_abc123").credentials
client.webauthn.delete_credential(credential_id)

WebAuthn API Endpoints

Method	Endpoint	Description
POST	/v1/webauthn/register/options	Generate passkey registration options
POST	/v1/webauthn/register/verify	Verify registration and store credential
POST	/v1/webauthn/enrollment-sessions	Issue a one-use hosted enrollment link (authenticated)
GET	/passkey-enroll	Hosted browser enrollment (feature-gated)
POST	/v1/webauthn/enroll/options	Ticket-bound hosted registration options
POST	/v1/webauthn/enroll/verify	Verify the passkey and consume the ticket
GET	/v1/webauthn/credentials	List WebAuthn credentials for a principal
DELETE	/v1/webauthn/credentials/:id	Delete a credential
POST	/v1/webauthn/assert/options	Generate assertion options for consent
POST	/v1/webauthn/assert/verify	Verify assertion during consent
PATCH	/v1/me	Update developer settings (FIDO config)

Hosted enrollment requires PASSKEY_ENROLLMENT_ENABLED=true and a correctly configured HTTPS FIDO_ORIGIN/FIDO_RP_ID; the server feature is off by default. Live-mode consent requires an existing passkey, with no weaker fallback. See the WebAuthn guide for the customer identity-binding and one-use link requirements. The guide also distinguishes the hosted SDK methods from the REST-only custom assertion ceremony.

For a production rollout, deploy the auth service and the /passkey-enroll hosting rewrite before enabling the flag. Then run Production Passkey and Irregularity E2E from GitHub Actions. That test creates an isolated live account, enrolls a virtual passkey, approves consent, checks portable evidence in an opt-in VC, refreshes and delegates the grant, and tests alert-only and revoke modes. It leaves the test account and audit records in production; do not run it against a customer account.

The workflow also tests sandbox/live consent parity, two authenticator devices, denial, credential removal, one-use ticket replay, and live OAuth approval and denial after interactive principal selection. Credential removal blocks new assertions. A dashboard browser regression covers login, enrollment-link issuance, registration, listing and confirmed removal using a bodyless DELETE; the portal sends JSON content type only for requests with a JSON body. Revoke existing grants separately. Agents with issued VCs cannot be hard-deleted (409 AGENT_HAS_CREDENTIAL_HISTORY); suspend them and revoke their grants to retain verifiable status history. Chromium virtual-authenticator checks are not certification of every physical device or browser.

</details> <details> <summary><strong>Verifiable Credentials</strong> — W3C VC-JWT issuance and verification</summary>

Verifiable Credentials

Grantex can issue W3C Verifiable Credentials (VCs) alongside standard JWTs. While JWTs are optimized for real-time authorization, VCs provide a portable, tamper-evident, standards-based proof of authorization that can be presented to any verifier — including systems outside the Grantex ecosystem.

Why VCs Matter for Agents

In agentic commerce, an agent acting on your behalf needs to prove its authorization to third-party services that may not integrate with Grantex directly. A Verifiable Credential is a self-contained, cryptographically signed document that any party can verify using the issuer's published DID document — no API calls, no accounts, no trust relationships required.

How It Works

When exchanging an authorization code for a grant token, pass credentialFormat: "vc-jwt" to receive a Verifiable Credential alongside the standard grant token:

const result = await grantex.tokens.exchange({
  code,
  agentId: agent.id,
  credentialFormat: 'vc-jwt',   // opt-in to VC issuance
});

console.log(result.grantToken);         // standard RS256 JWT (unchanged)
console.log(result.verifiableCredential); // W3C VC-JWT

result = client.tokens.exchange(ExchangeTokenParams(
    code=code,
    agent_id=agent.id,
    credential_format="vc-jwt",
))

print(result.verifiable_credential)  # W3C VC-JWT

Credential Types

Type	Description
AgentGrantCredential	Issued for direct grants — attests that a principal authorized an agent with specific scopes
DelegatedGrantCredential	Issued for delegated grants — includes the full delegation chain

Verifying a VC

const verification = await grantex.credentials.verify(vcJwt);
console.log(verification.valid);
console.log(verification.credentialSubject);
console.log(verification.issuer); // "did:web:grantex.dev"

verification = client.credentials.verify(vc_jwt)
print(verification.valid)
print(verification.credential_subject)

Revocation via StatusList2021

Grantex implements the W3C StatusList2021 revocation mechanism. Each credential references a status list entry. When a grant is revoked, the corresponding bit in the status list is flipped, and any verifier checking the credential sees it as revoked.

// Check a specific credential's status
const cred = await grantex.credentials.get(credentialId);
console.log(cred.status); // "active" or "revoked"

// List credentials with filters
const { credentials } = await grantex.credentials.list({
  grantId: 'grnt_01HXYZ...',
  status: 'active',
});

Portable Passkey Evidence

The hosted Grantex service enabled portable evidence on September 27, 2026 and passed the production passkey, delegation, refresh, and revocation E2E workflow. Self-hosted installations still default to off and must follow the rollout guide. The current published TypeScript, Python, and Go SDKs expose typed grant evidence references and VC attestations; custom assertion ceremonies still use the REST endpoints.

Live consent requires a passkey assertion. With PORTABLE_WEBAUTHN_EVIDENCE_ENABLED=true (source default off), newly issued grants and grant tokens carry a signed webauthnEvidence summary with the assertion digest. Activation also requires IRREGULARITY_CASCADE_REVOCATION_ENABLED=true and PORTABLE_WEBAUTHN_EVIDENCE_STATUS_CHECK_ENABLED=true, so automatic grant revocation updates delegated grants, VCs, and public status-list bits, and issuer verification checks the underlying grant. Keep both safety flags on if portable issuance is later rolled back. Request credentialFormat: 'vc-jwt' (or 'both') at token exchange to receive a signed VC with the raw assertion, public key, challenge, RP ID, origin, and prior counter in vc.evidence. With this rollout enabled or evidence already present, the VC is issued atomically with the grant. Delegated grants inherit this original-ceremony evidence, not a new human approval. Historical grants without captured evidence cannot be upgraded retroactively; the principal must complete a new consent ceremony after activation.

Verify the VC signature, expiry and a fresh revocation status list; require your trusted RP ID and origin; reverify the WebAuthn assertion; and compare its digest to the grant-token reference. Offline signature verification alone does not prove current grant status. The authenticator signs a challenge, not the grant's scopes. Grantex's issuer signature supplies that binding and attests its credential enrollment; a verifier must decide whether to trust that issuer and enrollment process. Raw evidence contains a stable credential public key and is opt-in because it can correlate presentations. No payment-network certification is implied. See the WebAuthn guide and VC guide.

DID Infrastructure

Grantex publishes a W3C DID document at /.well-known/did.json (did:web:grantex.dev). This document contains the public keys used to sign Verifiable Credentials, enabling any party to verify credentials without contacting Grantex:

curl https://api.grantex.dev/.well-known/did.json

Verifiable Credentials API Endpoints

Method	Endpoint	Description
GET	/v1/credentials/:id	Retrieve a Verifiable Credential
GET	/v1/credentials	List Verifiable Credentials
POST	/v1/credentials/verify	Verify a VC-JWT
GET	/v1/credentials/status/:id	StatusList2021 credential
GET	/.well-known/did.json	W3C DID document

</details> <details> <summary><strong>MPP Agent Identity</strong> — verifiable agent identity for machine payments</summary>

MPP Agent Identity

MPP (Machine Payments Protocol) defines HTTP 402 payment flows for software clients. Grantex can attach an AgentPassportCredential based on W3C Verifiable Credentials 2.0 so a configured merchant can verify agent, principal, category, amount-limit, expiry, and delegation claims after obtaining the issuer keys and current status data. The credential carries authorization context; it does not prove that a payment settled, an order was approved, or a provider accepted the transaction.

Why Agent Passports?

A wallet or payment-source identifier may not tell a merchant which internal agent is acting, which principal delegated authority, or which policy limits apply. An agent passport can supply that context when both sides integrate and enforce it.

The current Grantex passport shape uses Ed25519 signatures, a W3C VC 2.0 data model, configured purchase categories, transaction ceilings, expiry, delegation context, and StatusList2021-style status data. A relying merchant must still validate the issuer, refresh keys and status according to its risk policy, enforce the claims at the protected action, and run its own payment, fraud, sanctions, order, and settlement controls.

How It Works

Step	Who	What
1. Issue	Authorized application	Requests an AgentPassportCredential with configured categories, amount ceiling, delegation context, and expiry
2. Store	Agent host	Stores the credential as sensitive authorization material
3. Present	Agent host	Attaches the credential to a supported MPP request when the integration is configured
4. Verify	Merchant service	Verifies signature, issuer, expiry, category, amount, and sufficiently current status data
5. Decide	Merchant and payment systems	Apply merchant policy plus independent payment, order, provider, and settlement checks before execution
6. Record	Each participating system	Records the events it is configured to observe; credential verification alone is not an execution or settlement audit log

Verification boundary: cached keys can support local signature checks after retrieval. Current revocation requires refreshed status data, and cache policy determines how quickly a relying service observes a change.

Credential Structure

Field	Description
id	urn:grantex:passport:<ulid>
issuer	did:web:grantex.dev
credentialSubject.id	Agent DID (did:grantex:ag_...)
credentialSubject.humanPrincipal	DID of the authorizing human
credentialSubject.organizationDID	Org DID (did:web:<domain>)
credentialSubject.grantId	Links to underlying Grantex grant
credentialSubject.allowedMPPCategories	inference, compute, data, storage, search, media, delivery, browser, general
credentialSubject.maxTransactionAmount	{ amount, currency } ceiling per transaction
credentialSubject.delegationDepth	Inherited from grant delegation chain
credentialStatus	StatusList2021 revocation entry
proof	Ed25519Signature2020

Issue a Passport

TypeScript:

import { Grantex } from '@grantex/sdk';

const grantex = new Grantex({ apiKey: process.env.GRANTEX_API_KEY });

const passport = await grantex.passports.issue({
  agentId: 'ag_01HXYZ...',
  grantId: 'grnt_01HXYZ...',
  allowedMPPCategories: ['inference', 'compute'],
  maxTransactionAmount: { amount: 50, currency: 'USDC' },
  paymentRails: ['tempo'],
  expiresIn: '24h',
});
// passport.passportId   → "urn:grantex:passport:01HXYZ..."
// passport.encodedCredential → base64url for X-Grantex-Passport header

cURL:

curl -X POST https://api.grantex.dev/v1/passport/issue \
  -H "Authorization: Bearer $GRANTEX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agentId": "ag_01HXYZ...",
    "grantId": "grnt_01HXYZ...",
    "allowedMPPCategories": ["inference", "compute"],
    "maxTransactionAmount": { "amount": 50, "currency": "USDC" },
    "expiresIn": "24h"
  }'

Attach to MPP Requests

import { createMppPassportMiddleware } from '@grantex/mpp';

const middleware = createMppPassportMiddleware({ passport });
const enrichedRequest = await middleware(new Request(url, init));
// enrichedRequest now has X-Grantex-Passport header
const response = await fetch(enrichedRequest);

Verify on the Merchant Side

Standalone verification:

import { verifyPassport } from '@grantex/mpp';

const verified = await verifyPassport(encodedCredential, {
  requiredCategories: ['inference'],
  maxAmount: 10,
});
// verified.humanDID        → "did:grantex:user_alice"
// verified.organizationDID → "did:web:acme.com"
// verified.allowedCategories → ["inference", "compute"]
// verified.maxTransactionAmount → { amount: 50, currency: "USDC" }

Express middleware (one-liner):

import { requireAgentPassport } from '@grantex/mpp';

app.use('/api/inference', requireAgentPassport({
  requiredCategories: ['inference'],
  maxAmount: 10,
}));
// req.agentPassport is populated on valid requests
// 403 with typed error code on invalid requests

Trust Registry

Query any organization's verified trust level — no authentication required:

curl https://api.grantex.dev/v1/trust-registry/did:web:grantex.dev
# {"organizationDID":"did:web:grantex.dev","trustLevel":"soc2","domains":["grantex.dev"]}

import { lookupOrgTrust } from '@grantex/mpp';

const record = await lookupOrgTrust('did:web:acme.com');
// record.trustLevel → "verified" | "soc2" | "basic"
// record.verificationMethod → "dns-txt" | "manual" | "soc2"

Revocation

Revoking a passport updates server-side status immediately. Local verifiers observe the change only after an online revocation check or status-data refresh, so cached results can lag:

await grantex.passports.revoke('urn:grantex:passport:01HXYZ...');
// Revocation-aware verification rejects after checking refreshed status data

Error Codes

Code	HTTP	Description
PASSPORT_EXPIRED	403	Credential validUntil has passed
PASSPORT_REVOKED	403	StatusList2021 bit is set
INVALID_SIGNATURE	403	Signature verification failed
UNTRUSTED_ISSUER	403	Issuer DID not in trusted list
CATEGORY_MISMATCH	403	Categories don't cover required service
AMOUNT_EXCEEDED	403	Max amount below required threshold
MISSING_PASSPORT	403	No X-Grantex-Passport header
MALFORMED_CREDENTIAL	403	Invalid base64url or missing VC fields

MPP Agent Identity API Endpoints

Method	Endpoint	Auth	Description
POST	/v1/passport/issue	API key	Issue AgentPassportCredential
GET	/v1/passports	API key	List passports (filter by agentId, grantId, status)
GET	/v1/passport/:id	API key	Retrieve passport by ID
POST	/v1/passport/:id/revoke	API key	Revoke passport (StatusList2021)
GET	/v1/trust-registry/:orgDID	None	Look up org trust record (public)
GET	/v1/trust-registry	API key; admin key (ADMIN_API_KEY) when TRUST_REGISTRY_ADMIN_LISTING_ENFORCED=true	List all trust records, across developers (operator)

See packages/mpp/ for full package docs. Demo: grantex.dev/mpp-demo.

</details> <details> <summary><strong>SD-JWT Selective Disclosure</strong> — privacy-preserving credential presentation</summary>

SD-JWT Selective Disclosure

Grantex supports SD-JWT (Selective Disclosure JWT) for privacy-preserving credential presentation. While a standard VC-JWT reveals all claims to every verifier, SD-JWT lets the holder choose exactly which fields to disclose — keeping everything else hidden.

Why SD-JWT?

In agentic commerce, different verifiers need different levels of information. A payment processor needs to know the agent's scopes and budget, but not the principal's identity. A compliance auditor needs the principal and timestamps, but not the scopes. SD-JWT enables minimum-disclosure presentations that satisfy each verifier's requirements without over-sharing.

How It Works

When exchanging an authorization code, pass credentialFormat: "sd-jwt" to receive an SD-JWT credential:

const result = await grantex.tokens.exchange({
  code,
  agentId: agent.id,
  credentialFormat: 'sd-jwt',   // opt-in to SD-JWT issuance
});

console.log(result.grantToken);         // standard RS256 JWT (unchanged)
console.log(result.sdJwt);              // SD-JWT with selective disclosure

result = client.tokens.exchange(ExchangeTokenParams(
    code=code,
    agent_id=agent.id,
    credential_format="sd-jwt",
))

print(result.sd_jwt)  # SD-JWT with selective disclosure

Creating a Presentation

The holder selects which claims to disclose when presenting to a verifier:

const presentation = await grantex.credentials.present({
  sdJwt: result.sdJwt,
  disclosedClaims: ['scopes', 'agentId'],   // only reveal these fields
});

// Send presentation to the verifier — they see scopes and agentId,
// but principalId, developerId, grantId, etc. remain hidden

presentation = client.credentials.present(
    sd_jwt=result.sd_jwt,
    disclosed_claims=["scopes", "agent_id"],
)

SD-JWT Format

An SD-JWT consists of: <issuer-jwt>~<disclosure1>~<disclosure2>~...~

Each disclosure is a base64url-encoded JSON array [salt, claim-name, claim-value]. The verifier can only see claims for which a disclosure is provided.

Disclosable Claims

Claim	Description
principalId	The end-user who authorized the grant
developerId	The developer who owns the agent
scopes	The authorized scopes
agentId	The agent's DID
grantId	The grant record identifier
issuedAt	When the credential was issued
expiresAt	When the credential expires

SD-JWT API Endpoints

Method	Endpoint	Description
POST	/v1/token	Exchange code for grant token + SD-JWT (with credentialFormat: "sd-jwt")
POST	/v1/credentials/verify	Verify an SD-JWT presentation

</details> <details> <summary><strong>Budget Controls</strong> — per-agent spending limits with atomic debit</summary>

Budget Controls

Grantex provides per-grant budget controls that let developers cap how much an agent can spend. Budget allocations are enforced atomically — if a debit would exceed the remaining balance, it fails with a 402 INSUFFICIENT_BUDGET error. Threshold alerts fire at 50% and 80% consumption, and the remaining budget is embedded in grant tokens via the bdg JWT claim.

// Allocate a budget to a grant
const budget = await grantex.budgets.allocate({
  grantId: 'grnt_01HXYZ...',
  amount: 1000,
  currency: 'USD',
});

// Debit against the budget
const debit = await grantex.budgets.debit({
  grantId: 'grnt_01HXYZ...',
  amount: 42.50,
  description: 'Flight booking',
});
console.log(debit.remaining); // 957.50

// Check the current balance
const balance = await grantex.budgets.balance('grnt_01HXYZ...');
console.log(balance.remainingBudget);

// List all budget transactions for a grant
cons