diff --git a/src/gateway/mcp-http.runtime.ts b/src/gateway/mcp-http.runtime.ts
index b3f044b4fd9..c2959677b22 100644
--- a/src/gateway/mcp-http.runtime.ts
+++ b/src/gateway/mcp-http.runtime.ts
@@ -14,6 +14,7 @@ import { resolveGatewayScopedTools } from "./tool-resolution.js";
 // context and caches the expensive schema projection for short bursts of tool
 // list/call traffic from the same MCP client.
 const TOOL_CACHE_TTL_MS = 30_000;
+const TOOL_CACHE_MAX_ENTRIES = 256;
 const NATIVE_TOOL_EXCLUDE = new Set(["read", "write", "edit", "apply_patch", "exec", "process"]);
 
 type CachedScopedTools = {
@@ -72,6 +73,11 @@ export class McpLoopbackToolCache {
       params.senderIsOwner === true ? "owner" : "non-owner",
     ].join("\u0000");
     const now = Date.now();
+    for (const [key, entry] of this.#entries) {
+      if (now - entry.time >= TOOL_CACHE_TTL_MS) {
+        this.#entries.delete(key);
+      }
+    }
     const cached = this.#entries.get(cacheKey);
     // Config object identity is part of the cache contract so explicit gateway
     // reloads invalidate tool scope and schema without filesystem polling.
@@ -88,10 +94,12 @@ export class McpLoopbackToolCache {
       time: now,
     };
     this.#entries.set(cacheKey, nextEntry);
-    for (const [key, entry] of this.#entries) {
-      if (now - entry.time >= TOOL_CACHE_TTL_MS) {
-        this.#entries.delete(key);
+    while (this.#entries.size > TOOL_CACHE_MAX_ENTRIES) {
+      const oldestKey = this.#entries.keys().next().value;
+      if (oldestKey === undefined) {
+        break;
       }
+      this.#entries.delete(oldestKey);
     }
     return nextEntry;
   }
diff --git a/src/gateway/mcp-http.test.ts b/src/gateway/mcp-http.test.ts
index d32338d8818..95358619a4b 100644
--- a/src/gateway/mcp-http.test.ts
+++ b/src/gateway/mcp-http.test.ts
@@ -105,6 +105,7 @@ import {
   ensureMcpLoopbackServer,
   startMcpLoopbackServer,
 } from "./mcp-http.js";
+import { McpLoopbackToolCache } from "./mcp-http.runtime.js";
 
 let server: Awaited<ReturnType<typeof startMcpLoopbackServer>> | undefined;
 
@@ -576,6 +577,43 @@ describe("mcp loopback server", () => {
     expect(getScopedToolsCall(3).currentInboundAudio).toBe(true);
   });
 
+  it("caps loopback tool cache cardinality by evicting oldest contexts", () => {
+    const cache = new McpLoopbackToolCache();
+    const baseParams = {
+      accountId: undefined,
+      cfg: { session: { mainKey: "main" } } as never,
+      currentChannelId: "telegram:chat123",
+      currentInboundAudio: undefined,
+      currentMessageId: undefined,
+      currentThreadTs: "thread-1",
+      inboundEventKind: "room_event",
+      messageProvider: "telegram",
+      senderIsOwner: true,
+      sessionKey: "agent:main:telegram:group:chat123",
+      sourceReplyDeliveryMode: "message_tool_only",
+    } satisfies Parameters<McpLoopbackToolCache["resolve"]>[0];
+
+    for (let index = 0; index < 257; index += 1) {
+      cache.resolve({
+        ...baseParams,
+        currentMessageId: `message-${index}`,
+      });
+    }
+    expect(resolveGatewayScopedToolsMock).toHaveBeenCalledTimes(257);
+
+    cache.resolve({
+      ...baseParams,
+      currentMessageId: "message-0",
+    });
+    expect(resolveGatewayScopedToolsMock).toHaveBeenCalledTimes(258);
+
+    cache.resolve({
+      ...baseParams,
+      currentMessageId: "message-256",
+    });
+    expect(resolveGatewayScopedToolsMock).toHaveBeenCalledTimes(258);
+  });
+
   it("adds empty properties for object schemas that omit properties", async () => {
     resolveGatewayScopedToolsMock.mockReturnValue({
       agentId: "main",