Skip to content

visado: competitor/adjacent landscape (2026-11-18)

No project ships an AIMS-shaped composition (WIMSE WIT/WPT verification + agent delegation chain + Txn-Token minting + AuthZEN PEP) as one component. The pieces exist separately: zeroid (agent IdP/OBO), kontxt (Txn-Tokens behind agentgateway ext_authz), tokenetes (Txn-Tokens on K8s/SPIFFE, maintenance mode), agentgateway (JWT + CEL + an ext_authz-based AuthZEN PEP + RFC 8693).

The closest competitor is aramase/kontxt: Go, Txn-Token draft-08, an Envoy ext_authz adapter, and agentgateway as data plane. That is visado’s data path minus WIMSE/AIMS/AuthZEN. It has 26 stars, a single maintainer, and no push since 2026-06-12.

tokenetes is in maintenance mode, not archived. The main author has not committed since 2025-07. Since then the only activity is dependency/CI bumps (2026-07) by Atul Tulshibagwale, now at CrowdStrike.

Proxy foundation: no proxy of our own. tonic plus envoy-types for ext_authz (the primary mode) and axum for the token/AuthZEN HTTP endpoints. A standalone PEP, if one is needed later, would use hyper 1.x / hyper-util, not pingora.

Project Lang / License Stars / last push Specs claimed Overlap vs visado
tokenetes/tokenetes (ex-Tratteria) Go / Apache-2.0 66 / 2026-10-29 draft-ietf-oauth-transaction-tokens Txn-Token issuer only
highflame-ai/zeroid Go / Apache-2.0 213 / 2026-11-18 OAuth 2.1, RFC 8693/7523/9449/7591/9396/9068, CIBA, CIMD, ID-JAG, WIMSE/SPIFFE URIs, WIMSE proof tokens, SSF/CAE Agent IdP and delegation; no Txn-Token, no AuthZEN PEP
alibaba/open-agent-auth Java / Apache-2.0 92 / 2026-03-15 draft-liu-agent-operation-authorization, draft-ietf-wimse-workload-creds, OIDC, W3C VC, OPA WIT use; own token (AOAT)
Delego-Dev/delego Python / Apache-2.0 25 / 2026-06-11 own “delego wire spec” 0.3 Intent-bound action approval; not identity
microsoft/agent-governance-toolkit Python (+Rust core) / MIT 6,576 / 2026-11-17 Own RFC-2119 specs (AgentMesh Identity & Trust), SPIFFE/DID/mTLS, OPA/Cedar In-process SDK governance
Infisical/agent-vault Go / MIT + ee/ 2,515 / 2026-11-15 none (MITM credential proxy) Credential brokering; superseded by hosted product
kagenti -> rossoctl/rossoctl (AuthBridge) Python/Go / Apache-2.0 304 / 2026-11-14 RFC 8693, SPIFFE, Keycloak Sidecar JWT validation + exchange
agentgateway/agentgateway Rust / Apache-2.0 6,285 / 2026-11-18 (v1.7.0, 2026-10-27) JWT/OIDC, MCP auth, RFC 8693, ID-JAG/XAA, AuthZEN (example), Envoy ext_authz/ext_proc Data plane; visado’s host
IBM/mcp-context-forge Python / Apache-2.0 4,849 / 2026-11-18 OAuth, JWT, RFC 8693 token exchange, RBAC, plugins MCP gateway/registry
stacklok/toolhive Go / Apache-2.0 2,343 / 2026-11-18 OIDC, RFC 8693/OBO, Cedar authz, embedded auth server with SPIFFE + ID-JAG MCP runtime + authz
aramase/kontxt Go / Apache-2.0 26 / 2026-06-12 Txn-Tokens draft-08, RFC 8693, CEL Closest twin
zalando-incubator/agentic-identity-broker Go (+React UI) / MIT 80 / 2026-11-18 (v0.1.147, 2026-11-17) OAuth2 AS (proxy/local/hybrid), RFC 8693 exchange + impersonation, CIMD, DCR, PKCE, RFC 8707; CEL + OPA; Envoy ext_proc Same gateway slot (agentgateway sidecar); consent/vault/3rd-party token focus
  • Architecture: three parts: a TraT issuance service, a K8s controller (tconfigd) for config CRDs, and per-service verification sidecars (tokenetes-agent). A gateway exchanges the user access token for a TraT. Repo layout is service/, kubernetes/, docs/.
  • Hard-wired: CNCF/Palark both state that apps “must use SPIFFE for service-to-service trust”. Config flows through a K8s controller.
  • Activity:
    • v1.0.1 shipped 2025-07-21.
    • The top contributor, kchiranjewee63 (98 commits), last committed 2025-07.
    • The 2026 commits are a Go/crypto bump and a golangci v2 migration by tulshi (Atul Tulshibagwale, profile: “Senior Director, Continuous Identity Strategy at CrowdStrike”, Txn-Token co-author and SSF/AuthZEN co-chair).
    • tokenetes-agent: last push 2025-04, 3 stars.
  • Open issues (5): configurable sub claim (#81, 2026-05), scope→purp (#78), a dependabot PR, and #86 opened by tulshi (2026-10).
  • CNCF health: “Concerning (46)”.
  • SGNL: CrowdStrike announced the deal 2026-01-08 (~$740M) and closed it 2026-02-20.
  • Verdict (inference): not abandoned on paper, but a single-person, maintainer-of-last-resort project with no feature work since 2025. It risks a CNCF archival review.
  • What it is: a full agent IdP with Postgres-backed OAuth AS (the Postgres part is inferred from migrations/).
  • Tokens: the WIMSE URI is sub, with spiffe://{domain}/{account}/{project}/{type}/{id} identities.
  • Delegation: RFC 8693 OBO with scope attenuation, delegation_depth, and cascade revocation via SSF/CAE.
  • Other features: pluggable attestation, plus DPoP, CIMD and ID-JAG.
  • Activity: created 2026-03, 66 open issues.
  • vs visado: it issues identities; visado verifies and enforces. Complementary: zeroid could be an upstream issuer that visado validates. visado should not become an AS.
  • Three-layer binding: ID Token → WIT → AOAT.
  • Request-level virtual workloads with temporary keys.
  • W3C VC audit, OPA/RAM/ACL policy.
  • Spring/Java, public beta, quiet since 2026-03.
  • Spec: an individual draft (draft-liu-…), not the WG direction.
  • Deterministic intent-hash binding.
  • Single-use approvals bound to an action fingerprint.
  • Ed25519 hash-chained audit.
  • Rides credential brokers (explicitly Agent Vault).
  • Not identity. It is a possible AuthZEN PDP-side idea, not a competitor.
  • Pip-installable in-process governance: policy (YAML/OPA/Cedar via the “Agent Control Specification”, Rust core), identity “SPIFFE/DID/mTLS” (e.g. did:mesh:agent-1), trust scoring, audit, an MCP security gateway.
  • SDKs in five languages, many stars.
  • Own specs, no IETF Txn-Token/AuthZEN (code search: 0 AuthZEN hits).
  • vs visado: SDK-in-the-agent vs network PEP. Different layer.
  • MITM HTTP credential proxy: agents never hold secrets.
  • The README now says the OSS repo is “no longer recommended”, in favor of hosted Infisical Agent Vault.
  • License: MIT core + ee/.
  • Orthogonal: secret injection, not identity verification.
  • kagenti/kagenti redirects to rossoctl/rossoctl.
  • AuthBridge sidecar setup: an operator injects Envoy + a Go processor + spiffe-helper + Keycloak client registration.
  • Inbound plugin jwt-validation, outbound plugin token-exchange (RFC 8693, audience-scoped).
  • Maintainers are IBM Research. “Road to CNCF” is listed in the leads’ responsibilities.
  • vs visado:
    • Keycloak-centric exchange, no Txn-Token/WIMSE WIT verification.
    • Potential consumer of visado as its ext_authz.
    • Watch for CNCF positioning collision.

agentgateway (Linux Foundation, Solo.io-originated)

Section titled “agentgateway (Linux Foundation, Solo.io-originated)”
  • Security policies: JWT, basic, API key, OIDC browser auth, HTTP and network CEL authorization, MCP authentication (OAuth 2.0/JWT), MCP authorization (CEL over mcp.tool.name etc.), backend auth, CSRF, CORS, external authorization.
  • Code: http/{jwt,ext_authz,ext_proc,authorization,oauth}.rs, auth/oauth/cross_app_access.rs (ID-JAG).
  • Extension points:
    1. Envoy-compatible extAuthz, gRPC or HTTP. It supports includeRequestHeaders, a body, CEL metadata, includeResponseHeaders, conditional per-mcp.tool.
    2. extProc.
    3. ExtMcp, a new gRPC service “modeled on Envoy ext_authz, but at the MCP method layer”. CheckRequest gets method, params, headers and CEL metadata. It returns pass/mutated/error plus a HeaderMutation and a KV bag exposed as extMcp.<key> to CEL.
    4. CredentialProvider.FetchSecret(uri, actor_spiffe_id).
  • AuthZEN: examples/mcp-authzen already makes agentgateway an AuthZEN PEP (ext_authz HTTP mode with a CEL-built AuthZEN body, Keycloak as PDP) and does RFC 8693 oauthTokenExchange for backend auth.
  • No WIMSE, no Txn-Token (code search 0 hits).
  • Python MCP/A2A/REST gateway and registry.
  • Auth: basic, JWT (symmetric JWT_SECRET_KEY by default), OAuth with token-exchange service and cache, RBAC, plugin framework.
  • No AuthZEN. SPIFFE appears only in plugin docs.
  • MCP server runtime with per-request authn/authz.
  • pkg/auth has obo, upstreamswap, tokensource, awssts, dcr.
  • pkg/authz uses Cedar (plus an HTTP authorizer).
  • pkg/authserver is an embedded AS with SPIFFE bundle sources, ID-JAG, authorization_details and delegated-agent flows.
  • It is drifting toward the agent-IdP space.
  • Components: TTS (RFC 8693 → txntoken+jwt, tctx/rctx, CEL issuance rules), an ext_authz adapter with generation and verification modes, a controller, and a Go SDK.
  • CRDs: TxTokenConfig, TransactionType, ServiceTokenRequirement, TokenPolicy.
  • Integrations: agentgateway standalone or the istio-agentgateway GatewayClass.
  • Inference: the author is a K8s SIG-Auth maintainer at Microsoft. I did not verify the affiliation via his GitHub profile.
  • Hard-wired: req_wl = K8s SA, K8s-only.

zalando-incubator/agentic-identity-broker (high priority)

Section titled “zalando-incubator/agentic-identity-broker (high priority)”
  • Basics:
    • Go + React consent SPA, MIT, 80 stars.
    • Created 2026-07-09, pushed 2026-11-18, release v0.1.147 (2026-11-17). Fast cadence: 94 open issues, ADRs up to 047, spec-kit specs/.
  • Governance: MAINTAINERS lists only two core maintainers, both @zalando.de (Magnus Jungsbluth, Jan Brennenstuhl). They hold 207 of ~280 commits. In practice Zalando-only, in the zalando-incubator org.
  • What it is:
    • Consent-driven OAuth2 broker plus a token vault for third-party SaaS tokens (Google, GitHub, Databricks, Linear).
    • Tokens are envelope-encrypted with AWS KMS + DynamoDB; app data lives in Postgres.
    • User auth is delegated to a trusted reverse proxy (principal header/JWT). The broker itself does not authenticate users.
    • OAuth2 server modes: proxy (front an upstream AS), local (mint its own JWTs with managed asymmetric keys + JWKS), hybrid.
    • CIMD inbound and outbound (private_key_jwt), DCR, PKCE.
  • Delegation chain representation:
    • Not a token chain. Delegation is a stored Grant (principal → agent → permission sets across services, at most one active grant per user+agent, optional valid_until) plus a shared per-user-per-service session.
    • The subject token carries the user (sub) and agent (azp) via configurable CEL extraction.
    • Only “impersonation” (Feature 037/ADR 045) mints a broker JWT with a single-level act.iss/act.sub (privileged client such as a Slack bridge acting for a user).
    • No nested act, no agent→agent multi-hop.
  • Tokens minted:
    • Local JWT access tokens (agent_id, sub, scope, optional act).
    • JWE session tokens during authorization.
    • Otherwise it returns the third-party provider’s access token from the vault via RFC 8693.
  • Specs absent: ARCHITECTURE.md and the docs site have zero mentions of Txn-Token, WIMSE, SPIFFE, AuthZEN or ID-JAG (grep of the full ARCHITECTURE.md, plus the docs intro and delegation pages).
  • agentgateway integration:
    • A standalone extproc-token-exchange binary implements Envoy ext_proc (not ext_authz).
    • agentgateway does jwtAuth and puts the raw token + resource URI into trusted aib.tokenexchange metadata.
    • ExtProc calls the broker’s /oauth2/token (RFC 8693, gateway authenticated by client_assertion). It replaces Authorization: Bearer <exchanged> and caches by (subjectToken, resourceURI) with singleflight.
    • Three fail-closed gates:
      1. Broker CEL on the exchange.
      2. Embedded OPA in ExtProc on the request/MCP body.
      3. A human tool-approval gate (long-poll synced).
    • Re-auth is an MCP 403 insufficient_scope + elicitation challenge.
  • Overlap with visado:
    • Same deployment slot: a gRPC sidecar to agentgateway that validates the inbound JWT and swaps the outbound credential.
    • Both claim “delegation chains for on-behalf-of”.
    • Both do per-request policy (OPA/CEL vs AuthZEN).
    • Both do MCP tool-level gating.
  • Differences:
    • It is user-consent and SaaS-credential-vault centric (outbound to external OAuth providers). visado is workload-identity and intra-domain-call-chain centric.
    • Identity: it trusts bearer JWTs + a proxy header. visado verifies WIMSE WIT/WPT / SVID proof of possession.
    • Chains: its delegation is a DB grant, plus single act for impersonation. visado verifies a cryptographic multi-hop chain in tokens.
    • Output: it injects a third-party access token. visado mints a Txn-Token.
    • Policy: its PDP is embedded OPA/CEL. visado uses an external, standard AuthZEN PDP.
    • Protocol: ext_proc (streaming, body access) vs visado’s ext_authz (single Check).
    • It is a stateful app (Postgres, KMS, DynamoDB, UI). visado should stay near-stateless.
  • Positioning (inference): complementary. visado sits in front (verify agent workload + chain → AuthZEN → Txn-Token for internal hops). The broker sits at the egress to third-party SaaS. visado should not build a consent UI or a token vault. A possible interop demo: a visado-issued Txn-Token / verified chain as the broker’s subject token.
  • Sources: https://github.com/zalando-incubator/agentic-identity-broker (ARCHITECTURE.md, MAINTAINERS, README), https://agenticidentitybroker.dev/docs/introduction , https://agenticidentitybroker.dev/docs/concepts/delegation-and-consent
  • Aembit: Workload IAM extended to “IAM for Agentic AI”. GA April 2026. Blended Identity (agent+human), managed MCP Authorization Server (OAuth 2.1 / DCR), MCP Identity Gateway VM, secretless credential injection. CrowdStrike AIDR integration (2026-09).
  • Keycard: launched 2025-10 with $38M (a16z, Acrew, Boldstart). Founders ex-Snyk and Auth0 (Jared Hanson). Task-scoped short-lived tokens, policy at runtime, SDKs, “federated trust fabric”. Fortune Cyber 60 (2026-10).
  • Astrix: NHI discovery-first. The Agent Control Plane is pre-approved profiles + short-lived creds. Agent Policies (allow/flag/block) launched at RSAC 2026. Posture/inventory, not a protocol PEP.
  • Defakto (ex-SPIRL): SPIFFE/WIMSE-grounded NHI platform, Series B $30.75M (2025-10). Pieter Kasselman (Defakto) is lead author of draft-ietf-wimse-aims-01 (2026-11-02). They are the most likely commercial AIMS implementer. visado can be the vendor-neutral OSS reference, with Defakto invited as reviewers.
  • Service: implement envoy.service.auth.v3.Authorization/Check(CheckRequest) -> CheckResponse.
  • Input: CheckRequest.attributes (AttributeContext). Headers, path and the optional body live under attributes.request.http. Peer certificate/principal is in attributes.source. Route context_extensions and metadata_context are also present.
  • Allow and inject a Txn-Token: status=OK + ok_response:
    • headers: Txn-Token: <jwt> via HeaderValueOption. The docs say append defaults false = overwrite. In current core v3, use append_action: OVERWRITE_IF_EXISTS_OR_ADD.
    • headers_to_remove: the inbound Authorization / client-supplied Txn-Token and WIMSE headers, so credentials aren’t forwarded.
    • Also available: response_headers_to_add, query_parameters_to_set/remove.
  • Deny: denied_response{status, headers (e.g. WWW-Authenticate), body}.
  • Error: error_response, which honors failure_mode_allow.
  • Metadata: CheckResponse.dynamic_metadata lands under envoy.filters.http.ext_authz for downstream filters and logs. Put txn id / agent chain there.
  • Ordering: only one of denied/ok/error may be set.
  • Rust: envoy-types 0.7.7 (pre-compiled protos, tonic AuthorizationServer, CheckRequestExt/CheckResponseExt helpers, MSRV 1.88) + tonic 0.14.9.
  • Envoy config: set with_request_body if WPT/HTTP-sig verification needs the body digest.
  • Same ext_authz proto (it vendors ext_authz.proto), so one server covers Envoy, Istio and agentgateway.
  • An ExtMcp implementation can come later for MCP-method-aware decisions (tool name/args into the AuthZEN resource/action).
  • Its native AuthZEN example is a CEL-templated HTTP call. visado’s value is doing it with verified identity + chain + Txn-Token issuance in one hop.
  • Register in meshConfig.extensionProviders[].envoyExtAuthzGrpc{service,port}.
  • Attach with AuthorizationPolicy action: CUSTOM, provider.name. Only ext_authz is supported. Off-mesh authorizers need a ServiceEntry.
  • kontxt shows the istio-agentgateway GatewayClass path.
Crate Version Updated
tonic 0.14.9 2026-10
envoy-types 0.7.7 2026-09
hyper 1.12.0 2026-10
axum 0.8.9 2026-04
pingora 0.9.1 2026-11, 27.9k stars
rustls 0.23.46 2026-11
jsonwebtoken 12.0.0 2026-11
josekit 0.10.3 2025-05, OpenSSL-backed, stale
spiffe 0.19.0 2026-10, Workload API / SVIDs
  • JOSE: use jsonwebtoken 12 for JWS sign/verify, or a thin layer over aws-lc-rs/ring. josekit pulls OpenSSL and is stale.
  • HTTP Message Signatures (WIMSE WPT alt): I did not survey crate maturity. visado will probably implement RFC 9421 itself.
  1. Primary: no proxy. visado = tonic ext_authz server + axum HTTP surface (token endpoint, AuthZEN client, JWKS, health). Envoy, Istio and agentgateway all speak ext_authz, and agentgateway already is the Rust agent data plane (6.3k stars, LF). Competing with it would lose.
  2. If a standalone PEP mode is needed: hyper 1.x + hyper-util + tower + rustls.
    • Same Tokio/tower stack as tonic/axum, so ext_authz logic can be reused as a tower layer.
    • rustls is first-class.
    • Small dependency and audit surface for a security component.
  3. Avoid pingora for now.
    • Its own server lifecycle/runtime model doesn’t compose with tonic/axum.
    • rustls support is still labeled “experimental”; default builds want OpenSSL/BoringSSL (clang/perl).
    • It is built for high-throughput CDN-style proxying, which visado doesn’t need.
  4. axum is an app framework, not a proxy. Use it for endpoints only.

No other open, vendor-neutral Rust implementation found in this survey composes the AIMS stack end to end at the enforcement point. It verifies WIMSE workload credentials (WIT + WPT / HTTP-sig proof of possession) and SPIFFE SVIDs. It validates the agent delegation chain (RFC 8693 act nesting / ID-JAG from any issuer such as zeroid, Keycloak or Okta). It asks an AuthZEN PDP for the decision, and mints a draft-conformant Txn-Token carrying the verified chain and context. All of this happens in one ext_authz call that works in Envoy, Istio and agentgateway unchanged. Competitors each cover one slice:

  • tokenetes/kontxt: Txn-Tokens, K8s-only, no WIMSE/AuthZEN.
  • zeroid/toolhive authserver: issuance.
  • agentgateway: CEL policy + a templated AuthZEN example.
  • AGT: in-process SDK.
  • Aembit/Keycard/Defakto: closed SaaS.

visado should be the spec-tracking conformance reference (test vectors, interop with draft revisions), not another platform.

  • Not an OAuth AS / agent IdP: no client registration, DCR, CIBA, consent (zeroid, toolhive authserver, Keycloak). Mint only Txn-Tokens.
  • Not a data plane/gateway: no routing, MCP multiplexing, LLM proxying (agentgateway, context-forge, toolhive).
  • Not a policy engine: be the AuthZEN PEP/client. Ship at most a trivial test PDP (Cedar/OPA/AGT ACS live elsewhere).
  • Not a credential vault/broker, consent UI, or secret injector (Agent Vault, Zalando agentic-identity-broker, agentgateway CredentialProvider, Aembit). The Zalando broker owns user consent + third-party SaaS token vault behind agentgateway ext_proc. visado stays ext_authz + Txn-Token for internal hops.
  • Not a K8s operator/CRD suite in v1 (kontxt, tokenetes, rossoctl). Keep config file/xDS-free and portable. K8s comes later as packaging.
  • Not intent/HITL approval or prompt inspection (delego, AGT, Astrix posture).
  • Don’t fork tokenetes. Offer migration notes instead, and possibly co-maintainership or a conformance-suite collaboration with Atul Tulshibagwale / the kontxt author.
  • Star counts and dates are from GitHub API on 2026-11-18.
  • Commercial capability claims come mostly from vendor blogs and press.
  • kontxt author affiliation is inference.
  • Pingora rustls status is as stated in its current README.