src/mcplocal/src/audit/collector.ts

@@ -41,6 +41,11 @@ export class AuditCollector {
     this.sessionPrincipals.set(sessionId, { ...existing, tokenName: token.tokenName, tokenSha: token.tokenSha });
   }
 
+  /** Look up the McpToken SHA for a session. Returns undefined for non-HTTP-mode sessions. */
+  getSessionMcpTokenSha(sessionId: string): string | undefined {
+    return this.sessionPrincipals.get(sessionId)?.tokenSha;
+  }
+
   /** Queue an audit event. Auto-fills projectName, userName, tokenName, and tokenSha. */
   emit(event: Omit<AuditEvent, 'projectName'>): void {
     const enriched: AuditEvent = { ...event, projectName: this.projectName };

src/mcplocal/src/gate/session-gate.ts

@@ -3,6 +3,21 @@
  *
  * Tracks whether a session has gone through the prompt selection flow.
  * When gated, only begin_session is accessible. After ungating, all tools work.
+ *
+ * Per-token ungate cache:
+ * When the caller authenticated via an `McpToken` (HTTP-mode service agent),
+ * we also remember the ungate keyed on the token's SHA. Subsequent sessions
+ * from the same token automatically start ungated for a TTL window.
+ *
+ * Why: LiteLLM and similar MCP-proxying clients don't preserve the
+ * `mcp-session-id` header across chat completion calls, so every tool call
+ * lands on a fresh upstream session — which would otherwise be gated anew,
+ * forcing the agent into a begin_session loop. Keying on the token (which IS
+ * preserved, because it's in the Authorization header) gives us a stable
+ * identity that survives stateless proxies.
+ *
+ * Claude Code's stdio path keeps its session-id, so this code is a no-op for
+ * that case (session-id ungate still applies, token ungate is purely additive).
  */
 
 import type { PromptIndexEntry, TagMatchResult } from './tag-matcher.js';
@@ -14,15 +29,37 @@ export interface SessionState {
   briefing: string | null;
 }
 
+interface TokenUngateEntry {
+  tokenSha: string;
+  tags: string[];
+  ungatedAt: number;
+  retrievedPrompts: Set<string>;
+}
+
+/** Default TTL for per-token ungate cache (1 hour). Tunable via env for testing. */
+const DEFAULT_TOKEN_UNGATE_TTL_MS = Number(process.env['MCPLOCAL_TOKEN_UNGATE_TTL_MS']) || 60 * 60 * 1000;
+
 export class SessionGate {
   private sessions = new Map<string, SessionState>();
+  private tokenUngates = new Map<string, TokenUngateEntry>();
+  private readonly tokenUngateTtlMs: number;
 
-  /** Create a new session. Starts gated if the project is gated. */
+  constructor(tokenUngateTtlMs = DEFAULT_TOKEN_UNGATE_TTL_MS) {
-  createSession(sessionId: string, projectGated: boolean): void {
+    this.tokenUngateTtlMs = tokenUngateTtlMs;
+  }
+
+  /**
+   * Create a new session. Starts gated if the project is gated, UNLESS the
+   * caller's McpToken already ungated within the last TTL window — in which
+   * case the session inherits the previous tags + retrievedPrompts so the
+   * agent doesn't get the full gated greeting on every fresh session.
+   */
+  createSession(sessionId: string, projectGated: boolean, tokenSha?: string): void {
+    const priorEntry = tokenSha ? this.getActiveTokenEntry(tokenSha) : null;
     this.sessions.set(sessionId, {
-      gated: projectGated,
+      gated: projectGated && priorEntry === null,
-      tags: [],
+      tags: priorEntry ? [...priorEntry.tags] : [],
-      retrievedPrompts: new Set(),
+      retrievedPrompts: priorEntry ? new Set(priorEntry.retrievedPrompts) : new Set(),
       briefing: null,
     });
   }
@@ -37,18 +74,37 @@ export class SessionGate {
     return this.sessions.get(sessionId)?.gated ?? false;
   }
 
-  /** Ungate a session after prompt selection is complete. */
+  /** True when a token has an active (non-expired) ungate entry. */
-  ungate(sessionId: string, tags: string[], matchResult: TagMatchResult): void {
+  isTokenUngated(tokenSha: string): boolean {
+    return this.getActiveTokenEntry(tokenSha) !== null;
+  }
+
+  /**
+   * Ungate a session after prompt selection is complete.
+   *
+   * When `tokenSha` is supplied, also remember the ungate keyed on the token
+   * so future sessions from the same token start ungated (survives proxies
+   * that drop `mcp-session-id`).
+   */
+  ungate(sessionId: string, tags: string[], matchResult: TagMatchResult, tokenSha?: string): void {
     const session = this.sessions.get(sessionId);
     if (!session) return;
 
     session.gated = false;
     session.tags = [...session.tags, ...tags];
 
-    // Track which prompts have been sent
     for (const p of matchResult.fullContent) {
       session.retrievedPrompts.add(p.name);
     }
+
+    if (tokenSha !== undefined && tokenSha !== '') {
+      this.tokenUngates.set(tokenSha, {
+        tokenSha,
+        tags: [...session.tags],
+        ungatedAt: Date.now(),
+        retrievedPrompts: new Set(session.retrievedPrompts),
+      });
+    }
   }
 
   /** Record additional prompts retrieved via read_prompts. */
@@ -73,4 +129,19 @@ export class SessionGate {
   removeSession(sessionId: string): void {
     this.sessions.delete(sessionId);
   }
+
+  /** Forget a token's ungate entry (e.g. on revocation signal). */
+  revokeToken(tokenSha: string): void {
+    this.tokenUngates.delete(tokenSha);
+  }
+
+  private getActiveTokenEntry(tokenSha: string): TokenUngateEntry | null {
+    const entry = this.tokenUngates.get(tokenSha);
+    if (!entry) return null;
+    if (Date.now() - entry.ungatedAt > this.tokenUngateTtlMs) {
+      this.tokenUngates.delete(tokenSha);
+      return null;
+    }
+    return entry;
+  }
 }

src/mcplocal/src/proxymodel/plugin-context.ts

@@ -25,6 +25,13 @@ export interface PluginContextDeps {
   queueNotification: (notification: JsonRpcNotification) => void;
   postToMcpd: (path: string, body: Record<string, unknown>) => Promise<unknown>;
   auditCollector?: AuditCollector;
+  /**
+   * Resolves the principal's McpToken SHA for this session, if the caller
+   * authenticated via an McpToken. Called lazily so the value reflects the
+   * session's current state even when the token is attached after the plugin
+   * context is created.
+   */
+  getMcpTokenSha?: () => string | undefined;
 }
 
 /**
@@ -55,6 +62,11 @@ export class PluginContextImpl implements PluginSessionContext {
     this.deps = deps;
   }
 
+  /** McpToken SHA for the current caller, or undefined for STDIO/session-auth callers. */
+  getMcpTokenSha(): string | undefined {
+    return this.deps.getMcpTokenSha?.();
+  }
+
   registerTool(tool: ToolDefinition, handler: VirtualToolHandler): void {
     this.virtualTools.set(tool.name, { definition: tool, handler });
   }

src/mcplocal/src/proxymodel/plugin.ts

@@ -50,6 +50,14 @@ export interface PluginSessionContext {
 
   // Audit event emission (auto-fills sessionId and projectName)
   emitAuditEvent(event: Omit<AuditEvent, 'sessionId' | 'projectName'>): void;
+
+  /**
+   * McpToken SHA for the current caller, or undefined if the session was
+   * authenticated via a User session (STDIO/Claude Code path). Plugins can use
+   * this to key state on the token principal rather than the session-id —
+   * useful when the session-id doesn't survive a proxy (e.g. LiteLLM).
+   */
+  getMcpTokenSha(): string | undefined;
 }
 
 // ── Virtual Server ──────────────────────────────────────────────────

src/mcplocal/src/proxymodel/plugins/gate.ts

@@ -40,7 +40,11 @@ export function createGatePlugin(config: GatePluginConfig = {}): ProxyModelPlugi
     description: 'Gated session flow: begin_session → prompt selection → ungate.',
 
     async onSessionCreate(ctx) {
-      sessionGate.createSession(ctx.sessionId, isGated);
+      // Pass the caller's McpToken SHA so the gate can honor a cross-session
+      // ungate cache keyed on the token principal. Fixes the LiteLLM case where
+      // each tool call lands on a fresh mcp-session-id → would otherwise loop
+      // on begin_session forever.
+      sessionGate.createSession(ctx.sessionId, isGated, ctx.getMcpTokenSha());
 
       // Register begin_session virtual tool
       ctx.registerTool(getBeginSessionTool(llmSelector), async (args, callCtx) => {
@@ -264,8 +268,9 @@ async function handleBeginSession(
     matchResult = tagMatcher.match(tags, promptIndex);
   }
 
-  // Ungate the session
+  // Ungate the session (and remember the ungate per McpToken if this is a
-  sessionGate.ungate(ctx.sessionId, tags, matchResult);
+  // service-token request, so the next session from the same token skips the gate).
+  sessionGate.ungate(ctx.sessionId, tags, matchResult, ctx.getMcpTokenSha());
   ctx.queueNotification('notifications/tools/list_changed');
 
   // Audit: gate_decision for begin_session
@@ -451,8 +456,8 @@ async function handleGatedIntercept(
     const promptIndex = await ctx.fetchPromptIndex();
     const matchResult = tagMatcher.match(tags, promptIndex);
 
-    // Ungate the session
+    // Ungate the session (and remember per-token if the caller is a McpToken).
-    sessionGate.ungate(ctx.sessionId, tags, matchResult);
+    sessionGate.ungate(ctx.sessionId, tags, matchResult, ctx.getMcpTokenSha());
     ctx.queueNotification('notifications/tools/list_changed');
 
     // Audit: gate_decision for auto-intercept
@@ -522,7 +527,7 @@ async function handleGatedIntercept(
     return response;
   } catch {
     // If prompt retrieval fails, just ungate and route normally
-    sessionGate.ungate(ctx.sessionId, tags, { fullContent: [], indexOnly: [], remaining: [] });
+    sessionGate.ungate(ctx.sessionId, tags, { fullContent: [], indexOnly: [], remaining: [] }, ctx.getMcpTokenSha());
     ctx.queueNotification('notifications/tools/list_changed');
     return ctx.routeToUpstream(request);
   }

src/mcplocal/src/router.ts

@@ -198,6 +198,10 @@ export class McpRouter {
         return this.mcpdClient.post(path, body);
       },
       ...(this.auditCollector ? { auditCollector: this.auditCollector } : {}),
+      // Lazily resolve the caller's McpToken SHA via the audit collector's
+      // session principal map. The token is attached in onsessioninitialized,
+      // which runs before any plugin context is created, so this is stable.
+      getMcpTokenSha: () => this.auditCollector?.getSessionMcpTokenSha(sessionId),
     };
 
     ctx = new PluginContextImpl(deps);

src/mcplocal/tests/session-gate.test.ts

@@ -152,4 +152,76 @@ describe('SessionGate', () => {
     expect(gate.isGated('s1')).toBe(false);
     expect(gate.getSession('s2')!.tags).toEqual([]); // s2 untouched
   });
+
+  describe('per-McpToken ungate cache', () => {
+    it('new session from an already-ungated token starts ungated, with prior tags + prompts', () => {
+      const gate = new SessionGate();
+      gate.createSession('session-1', true, 'tokA');
+      expect(gate.isGated('session-1')).toBe(true);
+
+      gate.ungate('session-1', ['ops'], makeMatchResult(['runbook']), 'tokA');
+      expect(gate.isTokenUngated('tokA')).toBe(true);
+
+      // LiteLLM semantics: same token, brand-new session-id.
+      gate.createSession('session-2', true, 'tokA');
+      expect(gate.isGated('session-2')).toBe(false);
+      const s2 = gate.getSession('session-2')!;
+      expect(s2.tags).toContain('ops');
+      expect(s2.retrievedPrompts.has('runbook')).toBe(true);
+    });
+
+    it('does not persist across tokens', () => {
+      const gate = new SessionGate();
+      gate.createSession('s1', true, 'tokA');
+      gate.ungate('s1', ['ops'], makeMatchResult(['p']), 'tokA');
+
+      // Different token → fresh gated session.
+      gate.createSession('s2', true, 'tokB');
+      expect(gate.isGated('s2')).toBe(true);
+      expect(gate.isTokenUngated('tokB')).toBe(false);
+    });
+
+    it('is not triggered when no tokenSha is supplied (STDIO path)', () => {
+      const gate = new SessionGate();
+      gate.createSession('s1', true);
+      gate.ungate('s1', ['ops'], makeMatchResult(['p']));
+
+      // A second session with no token starts gated — STDIO semantics preserved.
+      gate.createSession('s2', true);
+      expect(gate.isGated('s2')).toBe(true);
+    });
+
+    it('honors the TTL window and expires', () => {
+      const gate = new SessionGate(50); // 50ms TTL for the test
+      gate.createSession('s1', true, 'tokA');
+      gate.ungate('s1', ['ops'], makeMatchResult(['p']), 'tokA');
+      expect(gate.isTokenUngated('tokA')).toBe(true);
+
+      return new Promise<void>((resolve) => setTimeout(() => {
+        expect(gate.isTokenUngated('tokA')).toBe(false);
+        gate.createSession('s2', true, 'tokA');
+        expect(gate.isGated('s2')).toBe(true);
+        resolve();
+      }, 70));
+    });
+
+    it('revokeToken clears the ungate entry immediately', () => {
+      const gate = new SessionGate();
+      gate.createSession('s1', true, 'tokA');
+      gate.ungate('s1', ['ops'], makeMatchResult(['p']), 'tokA');
+      expect(gate.isTokenUngated('tokA')).toBe(true);
+
+      gate.revokeToken('tokA');
+      expect(gate.isTokenUngated('tokA')).toBe(false);
+      gate.createSession('s2', true, 'tokA');
+      expect(gate.isGated('s2')).toBe(true);
+    });
+
+    it('empty-string tokenSha does not register an ungate entry', () => {
+      const gate = new SessionGate();
+      gate.createSession('s1', true, '');
+      gate.ungate('s1', ['ops'], makeMatchResult(['p']), '');
+      expect(gate.isTokenUngated('')).toBe(false);
+    });
+  });
 });