LangSmart Smartflow  |  Platform Brief  |  v1.3

Universal SSO Identity
Across the AI Stack

One authentication governs the LLM proxy, MCP tool servers, and A2A agents — sourced from your existing Active Directory, Entra ID, or any OIDC/SAML provider.
How It Works
Corporate Identity
AD / Entra ID / Okta
Kerberos / SAML / OIDC
POST /auth/sso
Smartflow JWT
sub · groups · dept · auth_method
LLM Proxy
Model allow-lists
Guardrails · Maestro
+
MCP Tool Servers
Group ACL
PKCE auto-inject
+
A2A Agents
Per-agent ACL
Caller in task metadata
The Problem Today

Most enterprise AI deployments treat authentication as solved at the edge — but identity evaporates the moment traffic enters the AI layer.

Shared service accounts hit every LLM endpoint. MCP tool servers issue tokens per-application with no link to the human operator. A2A agent tasks carry no caller identity. The result: no attribution, no per-user policy, and no audit trail.

When compliance requires knowing who submitted a prompt, what data they accessed through tools, and which agentacted on their behalf — current platforms have no answer.

The Smartflow Approach

Smartflow's SSO gateway exchanges any enterprise identity token — OIDC silent grant, Kerberos SPNEGO, SAML assertion, or trusted proxy header — for a single Smartflow JWT.

That JWT carries the user's AD group memberships, department, and authentication method. It is the only credential that flows through the entire stack. Users who are already logged into the corporate domain authenticate without a prompt.

Every downstream control — model allow-lists, MCP server ACLs, A2A agent ACLs, Maestro policy injection — reads from the same JWT. There is no second authentication, no credential synchronization, and no shared service account.

Identity Enforcement Across All Three Layers
LLM Proxy
Group-based guardrail policies. AD group membership drives model allow-lists, PII detection rules, and Maestro context injection. Every request is attributed to a real user in VAS logs.
MCP Tools
Per-server group ACL + PKCE auto-inject. Each MCP tool server declares allowed_groups and denied_groups. PKCE OAuth tokens are stored by JWT subject and injected automatically — users consent once, then all tool calls use their own credentials.
A2A Agents
Per-agent caller ACL + full audit trail. Agent profiles declare which groups may invoke them. The caller's subject, auth method, groups, and email are stored in every Task record. Wildcard patterns (Engineering-*) match sub-groups.
How Smartflow Compares
Capability Typical AI Gateway LiteLLM / Portkey Smartflow
User identity on LLM calls Service account only API key only Full SSO user identity
AD group-based model access Wildcard group patterns
MCP tool server group ACL Per-server allowed / denied
PKCE OAuth tied to user identity Auto-injected from JWT sub
A2A caller identity in task metadata Full audit trail per task
Zero re-login for SSO users Silent OIDC / Kerberos