# MudraID

> MudraID is action control for AI agents. Before an agent takes a consequential action — a refund, a payment, an admin change, a data export — MudraID answers three questions: Who is acting? Under whose authority? Should this exact action be allowed right now? It runs one loop: Authority → Decision → Containment → Proof.

Identity and gateway tools alone cannot answer all three questions at the moment an agent acts. MudraID connects them in one control loop:

- Authority: every agent and client has a verifiable cryptographic identity and a scoped, revocable grant traceable to a principal — no standing superuser.
- Decision: each request is mapped to the exact protected API or MCP tool being called and meets a default-deny check against current policy, trust, delegation, and revocation state.
- Containment: compromised authority is quarantined in one place, then revoked at the source and cascaded through delegations.
- Proof: allow and deny outcomes, reasons, and versions are captured as tenant-safe evidence a reviewer can independently check.

Agents authenticate as OAuth machine clients: the client authenticates at the OAuth token endpoint (`POST /oauth2/token`) with `private_key_jwt` (a signed assertion from a registered key) or, where approved, `client_secret_basic`, and receives a short-lived, scoped access token for an approved resource. The older native API-key-and-secret agent authentication is retired. MudraID is built on settled standards (OAuth 2.0, JWT, JWKS, OIDC) and aligned with emerging agent-identity standards (WIMSE/SPIFFE, MCP, A2A).

## Core concepts
- [What is AI agent identity](https://mudraid.ai/articles/what-is-ai-agent-identity.html): Why autonomous agents need a verifiable identity layer of their own.
- [Delegated authority for agents](https://mudraid.ai/articles/delegated-authority-for-agents.html): Scoped, time-bound, revocable permissions that let an agent act on behalf of a user or organization — without sharing keys.
- [Verifiable agents vs. bot detection](https://mudraid.ai/articles/verifiable-agents-vs-bot-detection.html): Why identity-based verification beats behavioral bot detection when legitimate agents need to act.
- [WIMSE and AIMS standards](https://mudraid.ai/articles/wimse-aims-standards.html): The open standards for agent and workload identity, and why to build on them.
- [Revocation and audit for autonomous systems](https://mudraid.ai/articles/revocation-and-audit.html): Revocation at the source, policy controls, and tamper-evident audit trails.
- [Agent-to-website verification at the edge](https://mudraid.ai/articles/agent-to-website-at-the-edge.html): Verifying agent requests at the CDN edge, API, or gateway without re-architecting.

## Guides
- [AI agent security](https://mudraid.ai/ai-agent-security.html): The flagship guide to AI agent security — the OWASP Top 10 for Agentic Applications, the confused-deputy problem, and why verifiable identity and delegated authority are the missing control plane.
- [AI agent identity: the complete guide](https://mudraid.ai/ai-agent-identity.html): The definitive overview of agent identity, the identity lifecycle, registry, ownership, public keys, key rotation, trust levels, delegation, verification, enforcement, governance, and audit.
- [AI agent authentication](https://mudraid.ai/ai-agent-authentication.html): JWTs, JWKS, and real-time verification for AI agents — how it relates to OAuth, and how it applies at the API, the edge, and MCP tools.
- [Build vs. buy: should you build AI agent identity yourself?](https://mudraid.ai/articles/build-vs-buy-agent-identity.html): An honest comparison of building an agent-identity layer in-house versus using a platform.

## Reference
- [AI agent identity glossary](https://mudraid.ai/glossary.html): Plain-language definitions of the core terms — agent identity, delegation, scope, verification, trust scoring, revocation, audit, JWT, JWKS, OAuth2, OIDC, WIMSE, SPIFFE, MCP, and A2A.
- [Full site content for LLMs](https://mudraid.ai/llms-full.txt): The key MudraID content in one plain-text file for ingestion.

## Product
- [Features](https://mudraid.ai/feature.html): Agent identity, delegation, authorization, verification, trust scoring, enforcement, governance, audit trail, key management, onboarding, accounts & orgs, standards, and the Python Agent SDK and FastAPI middleware (`mudraid-sdk` and `mudraid-platform-middleware` on PyPI).
- [Contact](https://mudraid.ai/contact.html): Get in touch or request access. Production access is by request; requests are reviewed.

## Trust
- [Security](https://mudraid.ai/security.html): MudraID's security approach and how to report a vulnerability.
- [Responsible disclosure](https://mudraid.ai/responsible-disclosure.html): Vulnerability disclosure policy, safe harbor, and scope.
- [Subprocessors](https://mudraid.ai/subprocessors.html): The vendors MudraID uses to deliver its services.
- [Data Processing Addendum](https://mudraid.ai/data-processing-addendum.html): How to request MudraID's DPA.
- [Privacy Policy](https://mudraid.ai/privacy.html): How MudraID collects, uses, and protects personal data.
- [Terms & Conditions](https://mudraid.ai/terms.html): The agreement governing use of MudraID's services.

## Documentation
Developer and decision-maker documentation, hosted at https://docs.mudraid.ai.
- [Introduction](https://docs.mudraid.ai/overview/introduction): What MudraID is and the trust problem it solves.
- [How it works](https://docs.mudraid.ai/overview/how-it-works): The end-to-end flow — onboarding, identity, delegation, action-time decisions, containment, and revocation.
- [Use cases](https://docs.mudraid.ai/overview/use-cases): Where action control for agents applies.
- [Quickstart: agent](https://docs.mudraid.ai/guides/get-started/quickstart-agent): Register an AI agent and authenticate it as an OAuth machine client.
- [Quickstart: platform](https://docs.mudraid.ai/guides/get-started/quickstart-platform): Onboard a platform and protect its APIs and MCP tools.
- [Trust model](https://docs.mudraid.ai/guides/concepts/trust-model): How MudraID establishes and evaluates trust.
- [Agent identity](https://docs.mudraid.ai/guides/concepts/agent-identity): The identity model for agents.
- [Auth tokens](https://docs.mudraid.ai/guides/concepts/auth-tokens): Token format, claims, and verification.
- [Platforms & scopes](https://docs.mudraid.ai/guides/concepts/platforms-scopes): Defining and enforcing scoped permissions.
- [Python Agent SDK](https://docs.mudraid.ai/sdks/python-sdk/overview): Authenticates an agent through a linked OAuth machine client (`private_key_jwt`, or approved `client_secret_basic`) and handles token caching and refresh. On PyPI as [mudraid-sdk](https://pypi.org/project/mudraid-sdk/), version 2.1.0 — `pip install mudraid-sdk`. Native API-key/secret authentication is retired.
- [FastAPI middleware](https://docs.mudraid.ai/sdks/middleware/overview): Route- and scope-gating for FastAPI/Starlette APIs. On PyPI as [mudraid-platform-middleware](https://pypi.org/project/mudraid-platform-middleware/), version 1.4.0 — `pip install mudraid-platform-middleware`.
- Node MCP adapter and sidecar: [@mudraid/adapter-node](https://www.npmjs.com/package/@mudraid/adapter-node) and [@mudraid/sidecar](https://www.npmjs.com/package/@mudraid/sidecar), version 1.1.2 on npm (alpha) — a decision core and a customer-hosted reverse proxy that require a signed live decision for protected MCP Streamable HTTP tool calls.
- Kong plugin: `mudraid-enforce` 1.1.0 is a prerelease and is not yet published to LuaRocks; install it from the rockspec in the public repository.
- [API reference](https://docs.mudraid.ai/api-reference/introduction): Public API contracts and versioning.
- [Glossary](https://docs.mudraid.ai/resources/glossary): Key terms.
- [FAQ](https://docs.mudraid.ai/resources/faq): Common questions about MudraID.

## Standards MudraID speaks
- Built on JWT, JWKS, RS256, OAuth2 Bearer, and OIDC Discovery.
- MCP (OAuth 2.1 + PKCE) authorization profile for tool-calling agents.
- A2A signed Agent Cards for agent-to-agent interoperability.
- OpenTelemetry and W3C Trace Context for end-to-end observability.
- Aligned with WIMSE (IETF workload identity) and SPIFFE.