Your Authz system shouldn't be allucinated
The compile-time safe auth engine for Rust. Enforce interactive human passkeys, MCP step-up protocols, and tamper-proof canonical digests - without running a sidecar.
cargo add authkestraGive AI agents tools, not secrets
authkestra-mcp puts every MCP tool call behind one of three security tiers.
Public
Read-only tools anyone can call. Caller identity is attached when present.
Scoped
Needs an OAuth 2.1 token with the right scopes, sender-bound with DPoP or mTLS.
Human-approved
A person approves the exact arguments with a passkey before the tool runs.
Arguments can’t drift
Approval is tied to a hash of the exact arguments. Change one value and the call fails.
Secrets never reach the model
API keys and refresh tokens stay in an encrypted server-side vault.
Miswiring fails to compile
A missing store or validator is a build error, not a production panic.
Anatomy of a guarded tool call
Follow one high-stakes tool call from request to execution.
- 01AI Agent CallMCP JSON-RPC
- 02Policy & JCS HashRFC 8785 SHA-256
- 03Passkey Step-UpWebAuthn Proof
- 04Scoped Vault TokenSingle-Use Proof
Intercept the call
The request is parsed and tied to the caller’s identity.
Canonicalize and check policy
The arguments are hashed (RFC 8785, SHA-256) and Cedar policy decides if the caller may proceed.
Human approves the exact digest
A person approves exactly those arguments with a passkey.
Run with a short-lived token
The server spends the one-time approval and issues a short-lived token. The model never sees the real credential.
Quickstart
Pick a scenario. Each one is a few lines of Rust.
Model Context Protocol tool security
Every MCP tool call goes through a security tier. High-stakes tools need a human passkey approval over the exact arguments.
Read the AI Agent & MCP guideuse authkestra::core::auth::ReproofRequirement;use authkestra::mcp::{McpRouter, McpSecurityTier, ToolDefinition}; let router = McpRouter::builder(pending_store, execution_store) .tool(ToolDefinition::new( "execute_transfer", schema, McpSecurityTier::human_approved( ["payments:write"], ReproofRequirement::new(300).require_step_up(true), 300, ), transfer_handler, )) .build();FIDO2 Passkeys with Re-Proof Gating
Passkeys with re-proof: adding a new authenticator means proving an existing one first, so a stolen session can’t take over the account.
Read the Passkeys guideuse authkestra::core::auth::ReproofRequirement;use authkestra::webauthn::WebAuthnAuthMethod; let passkeys = WebAuthnAuthMethod::new(webauthn, credential_store); // Re-proof enrolment gating prevents session-hijack escalationlet (challenge, enrolment) = passkeys.start_register( &caller_identity, &ReproofRequirement::new(300).require_step_up(true), "alice",)?;Stateless OAuth2 login
OAuth state and PKCE live in an encrypted cookie, so there is no session database to run.
Read the OAuth2 guideuse authkestra::Authkestra;use authkestra::axum::{AxumState, AxumStatelessExt};use authkestra::core::OAuth2Flow;use authkestra::providers::github::GithubProvider;use axum::Router;use tower_cookies::CookieManagerLayer; // requires tower-cookies let github = GithubProvider::new(client_id, client_secret, redirect_uri);let engine = Authkestra::builder() .provider(OAuth2Flow::new(github)) .jwt_secret(&token_secret) .build(); let app = Router::new() .merge(engine.axum_router_stateless()) .layer(CookieManagerLayer::new()) .with_state(AxumState::from(engine));What’s in the box
One engine, built from modules you can mix.
Sign-in methods
Server roles
authkestra-mcpGuard MCP tools for AI agentsauthkestra-opRun an OpenID Provider in-processauthkestra-resourceVerify DPoP and mTLS-bound tokensauthkestra-policyCedar access policies in microsecondsauthkestra-ssfRevoke sessions in real timeauthkestra-sessionToken rotation on Postgres, MySQL, SQLite, Redis
Where it stands
Version 0.15, so APIs can still change before 1.0. Built on audited crates such as ring, webauthn-rs and cedar-policy. A third-party audit is planned before 1.0. Security policy
Is it for you?
Pick it for Rust APIs, web apps and MCP servers that want auth in-process. Pick Keycloak, Zitadel or Ory if you need an admin console, SAML, or workforce SSO. You can also connect them over OIDC.
Frequently Asked Questions
Short answers. The docs have the rest.
How does it keep AI agents from misusing tools?
Every MCP tool sits in one of three tiers: public, scoped to an OAuth 2.1 token, or human-approved with a passkey. Approval covers a hash of the exact arguments, so a changed value fails the call. Third-party credentials stay in a server-side vault.
Should I use this or Keycloak, Ory or Zitadel?
Use Authkestra when you are building in Rust and want auth in-process with no sidecar. Use a standalone IdP if you need an admin console, SAML, or workforce SSO. You can connect the two over OIDC.
How does stateless OAuth work without a session database?
OAuth state and PKCE verifiers live in an AES-256-GCM encrypted cookie. On callback the engine decrypts it in memory, so you can scale out without Redis or SQL for sessions.
Does it work with my framework and database?
Yes. The core has no framework dependency. There are Axum and Actix-web adapters, and stores for memory, Redis, PostgreSQL, MySQL and SQLite. You can implement the store traits for anything else.