# Grantex

> Last updated: 2026-04-05

> Delegated authorization protocol for AI agents. What OAuth 2.0 is to humans, Grantex is to agents.

Grantex is an open protocol (Apache 2.0) that lets AI agents request scoped, human-approved, revocable permissions via signed JWTs. Any service can verify tokens offline using published JWKS — no runtime dependency on Grantex infrastructure.

## Quick Facts

- Website: https://grantex.dev
- Docs: https://docs.grantex.dev
- GitHub: https://github.com/mishrasanjeev/grantex
- Spec: https://github.com/mishrasanjeev/grantex/blob/main/SPEC.md
- IETF Draft: https://datatracker.ietf.org/doc/draft-mishra-oauth-agent-grants/
- License: Apache 2.0
- Protocol Version: 1.0 (Final, frozen)

## SDKs

- TypeScript: `npm install @grantex/sdk`
- Python: `pip install grantex`
- Go: `go get github.com/mishrasanjeev/grantex-go`
- CLI: `npm install -g @grantex/cli`

## Framework Integrations

- Anthropic SDK: `npm install @grantex/anthropic @grantex/sdk`
- LangChain: `npm install @grantex/langchain`
- CrewAI: `pip install grantex-crewai`
- OpenAI Agents SDK: `pip install grantex-openai-agents`
- Google ADK: `pip install grantex-adk`
- Vercel AI SDK: `npm install @grantex/vercel-ai`
- AutoGen: `npm install @grantex/autogen`
- MCP (Claude Desktop, Cursor, Windsurf): `npm install @grantex/mcp`
- Express.js middleware: `npm install @grantex/express`
- FastAPI middleware: `pip install grantex-fastapi`
- Reverse-proxy gateway: `npm install @grantex/gateway`

## Key Packages (New in v2.5)
- @grantex/gemma: Offline authorization for Gemma 4 on-device agents
- @grantex/mcp-auth: OAuth 2.1 auth server for MCP servers
- @grantex/dpdp: DPDP Act 2023 + EU AI Act compliance module

## Key Features

- Scoped permissions: `resource:action[:constraint]` format (e.g., `payments:initiate:max_500`)
- Human consent flow with plain-language UI
- Offline token verification via JWKS (RS256 JWT)
- Multi-agent delegation chains with scope narrowing
- Real-time revocation (single grant or cascade)
- Hash-chained audit trail
- FIDO2/WebAuthn passkey-based consent
- W3C Verifiable Credentials (VC-JWT) issuance
- SD-JWT selective disclosure
- DID infrastructure (did:web:grantex.dev)
- MPP Agent Passport for machine payments
- x402 agent spend authorization
- Budget controls with atomic debit
- Event streaming (SSE + WebSocket)
- Policy-as-code (OPA, Cedar)
- Usage metering and custom domains
- Terraform provider (5 resources, 2 data sources)
- Conformance test suite

## How It Works (5 Steps)

1. Developer registers an agent → gets a DID (`did:grantex:ag_...`)
2. Agent requests authorization for specific scopes → user sees consent UI
3. User approves → authorization code returned via redirect
4. Code exchanged for signed grant token (RS256 JWT)
5. Any service verifies the token offline via JWKS

## Compliance Coverage

- OWASP Agentic Top 10 (Dec 2025): ASI-01 (goal hijacking), ASI-03 (identity abuse), ASI-05 (privilege escalation), ASI-10 (rogue agents)
- EU AI Act (binding Aug 2026): Art. 9 (risk management), Art. 13 (transparency), Art. 14 (human oversight)
- NIST AI RMF + EO 14110: Govern 1.1 (accountability), Map 5.1 (attribution), Measure 2.5 (audit trails)
- SOC 2 Type I certified
- Full compliance matrix: https://docs.grantex.dev/guides/compliance-matrix

## Scope Enforcement (v0.3.4)

- `grantex.enforce(grant_token, connector, tool)` — verify JWT + check tool permission via manifest
- `grantex.load_manifest(manifest)` / `load_manifests([...])` — load tool permission definitions
- `grantex.wrap_tool(langchain_tool, connector, tool, grant_token)` — auto-enforce on LangChain tools
- `grantex.enforceMiddleware(options)` — Express/Fastify middleware for scope enforcement
- FastAPI: `from grantex.fastapi import GrantexEnforcer` — dependency injection for scope enforcement
- `enforce_mode="permissive"` — log-only mode without blocking (for migration)
- `ToolManifest(connector, tools)` — define custom connector permissions
- `Permission.READ/WRITE/DELETE/ADMIN` — hierarchy: admin > delete > write > read
- 54 pre-built manifests: from grantex.manifests.salesforce import manifest
- CLI: grantex manifest generate ./path — auto-generate manifests from connector code
- CLI: grantex manifest list — browse bundled manifests
- CLI: grantex enforce test --token X --connector Y --tool Z — dry-run
- Unknown tools denied by default (fail closed)
- Custom manifests: inline, JSON file, or extend pre-built
- Case study: AgenticOrg — 35 agents, 54 connectors, 340+ tools, all enforced

## Comparisons

- vs OAuth 2.0: Grantex adds cryptographic agent identity, delegation chains, action-level audit, and sub-second revocation
- vs API keys: Grantex provides per-agent scoping, user consent, delegation, and revocation granularity
- vs MCP Auth: Grantex works across all frameworks, not just MCP; adds VCs, FIDO2, budgets, and delegation

## Documentation

- Full docs: https://docs.grantex.dev/llms-full.txt