diff --git a/server/src/copilot.ts b/server/src/copilot.ts
index 766ce4f2..8582a85d 100644
--- a/server/src/copilot.ts
+++ b/server/src/copilot.ts
@@ -8,6 +8,7 @@ import {
 import { createCopilotHonoHandler } from "@copilotkit/runtime/v2/hono";
 import { z } from "zod";
 import { COMPUTER_GUIDANCE } from "../../shared/bot-prompt";
+import { grantedToolGuidance } from "./plugins/tools";
 import type { AgentActor } from "./agents/profile-types";
 import type { StallGuard } from "./channels/stall-guard";
 import type { DeploymentConfig } from "./config";
@@ -200,9 +201,18 @@ export function builtInAgentConfiguration(
 
   return {
     model: `${model.provider}/${model.defaultModel}`,
-    prompt: computerGuidance
-      ? `${agent.systemPrompt}\n\n${computerGuidance}`
-      : agent.systemPrompt,
+    /*
+     * The package's role, then what this Bot actually holds, then the computer.
+     *
+     * The grants go BEFORE the computer prose on purpose. That prose is long and emphatic about the
+     * browser and mentions connectors nowhere, so a Bot that read it last reached for the browser
+     * even when it held a tool for the exact system being asked about.
+     */
+    prompt: [
+      agent.systemPrompt,
+      ...(grantedToolGuidance(tools) ? [grantedToolGuidance(tools)] : []),
+      ...(computerGuidance ? [computerGuidance] : []),
+    ].join("\n\n"),
     apiKey,
     /*
      * A run stops after one step unless told otherwise, which for a Bot with tools means it calls
@@ -327,13 +337,37 @@ function remoteAgentWithStandingRole(
       ? { fetch: stallGuard.watch({ id: agent.id, name: agent.name }) }
       : {}),
   });
+  /*
+   * What this Bot holds, as a second standing message.
+   *
+   * Beside the role rather than inside it, because the role comes from the package and this comes
+   * from the grants: they change for different reasons and at different times. Sent on every run for
+   * the same reason the tools are, so switching a connector on reaches the next run.
+   *
+   * The remote path needs this more than the built-in one, not less. A framework Bot is handed the
+   * tools as an offer and decides for itself what to call, with `COMPUTER_GUIDANCE` as its whole
+   * prompt — a page about the browser that mentions connectors nowhere. That is the Bot that browsed
+   * to drive.google.com holding four Drive tools.
+   */
+  const holdings = grantedToolGuidance(tools);
+  const holdingsMessage = holdings
+    ? {
+        id: `granted-tools:${agent.id}`,
+        role: "system" as const,
+        content: holdings,
+      }
+    : null;
+
   remote.use((input, next) =>
     next.run({
       ...input,
       messages: [
         agent.standingMessage,
+        ...(holdingsMessage ? [holdingsMessage] : []),
         ...input.messages.filter(
-          (message) => message.id !== agent.standingMessage.id,
+          (message) =>
+            message.id !== agent.standingMessage.id &&
+            message.id !== holdingsMessage?.id,
         ),
       ],
       /*
diff --git a/server/src/plugins/tools.ts b/server/src/plugins/tools.ts
index 0bc35518..35aa4753 100644
--- a/server/src/plugins/tools.ts
+++ b/server/src/plugins/tools.ts
@@ -48,6 +48,48 @@ export function parametersFor(inputSchema: Record<string, unknown>): z.ZodType {
   return z.object({}).catchall(z.unknown());
 }
 
+/**
+ * What this Bot holds, said in its instructions rather than left to be inferred from a tool list.
+ *
+ * A tool array tells a model a tool exists. It does not tell it that the tool is the right way to
+ * reach that system, and it competes with a page of prose about the browser that every Bot is given
+ * whether or not it has any connectors at all. The browser prose wins: it is emphatic, it is about
+ * capability, and it says "never claim you cannot browse".
+ *
+ * So a Bot holding four Google Drive tools browsed to drive.google.com, met a sign-in page its
+ * container could never satisfy, and asked its person to sign in to a vendor that person had already
+ * connected. The tools were there the whole time.
+ *
+ * Generated from the grants rather than written down, because the point is that it tracks them. An
+ * administrator switching a connector on, or granting one more of its tools, changes what the Bot is
+ * told on its next run with nothing else to remember and nothing to keep in step.
+ *
+ * Empty when the Bot holds nothing, so a deployment with no connectors says nothing about them.
+ */
+export function grantedToolGuidance(tools: GrantedTool[]): string {
+  if (tools.length === 0) return "";
+
+  const bySystem = new Map<string, string[]>();
+  for (const tool of tools) {
+    // `mcp__server__tool`, which is the shape the model is offered.
+    const parts = tool.name.replace(/^mcp__/, "").split("__");
+    const system = parts.length > 1 ? (parts[0] as string) : "this deployment";
+    const rest = parts.length > 1 ? parts.slice(1).join("__") : tool.name;
+    bySystem.set(system, [...(bySystem.get(system) ?? []), rest]);
+  }
+
+  return [
+    "You can reach these systems directly, as the person asking, with their own access:",
+    ...[...bySystem.entries()].map(
+      ([system, names]) => `- ${system}: ${names.join(", ")}`,
+    ),
+    "Use them for anything about those systems. Do NOT browse to one of their websites instead: your",
+    "browser is signed in as nobody, so it sees less than these tools do and will meet a sign-in wall",
+    "that connecting an account has already solved. If one of these systems is involved but no tool",
+    "above covers the part you need, say which part is missing rather than going around it.",
+  ].join("\n");
+}
+
 /**
  * Every MCP tool granted to one Bot, ready to hand to the runtime.
  *
diff --git a/server/tests/copilot.test.ts b/server/tests/copilot.test.ts
index 5bd14d47..ee623640 100644
--- a/server/tests/copilot.test.ts
+++ b/server/tests/copilot.test.ts
@@ -9,6 +9,7 @@ import {
   resolveRuntimeAgents,
   standingRoleMessage,
 } from "../src/copilot";
+import { grantedToolGuidance } from "../src/plugins/tools";
 
 // Every agent row now joins its profile, so the row a coworker is built from always names it.
 const assistantRow = {
@@ -502,3 +503,63 @@ function fakeAgUiEndpoint() {
     [Symbol.asyncDispose]: () => server.stop(true),
   };
 }
+
+/**
+ * A Bot is told what it holds, not only handed it.
+ *
+ * A tool array tells a model a tool exists. It does not say the tool is the right way to reach that
+ * system, and it competes with `COMPUTER_GUIDANCE`: a page of emphatic prose about the browser that
+ * every Bot gets whether or not it has a single connector, and that mentions connectors nowhere.
+ *
+ * The browser prose won. A Bot holding four Google Drive tools browsed to drive.google.com, met a
+ * sign-in page its container could never satisfy, and asked its person to sign in to a vendor that
+ * person had already connected. Asked a question with no tool for it, another went reading a
+ * government website and looped on its 404 page.
+ *
+ * Both kinds are asserted because they are built by different functions, and the remote one is the
+ * one that failed in the product.
+ */
+describe("what a Bot is told it holds", () => {
+  const drive = [
+    { name: "mcp__google-drive__search_files" },
+    { name: "mcp__google-drive__read_file_content" },
+  ] as never[];
+
+  test("names the system and its tools", () => {
+    const guidance = grantedToolGuidance(drive);
+    expect(guidance).toContain("google-drive");
+    expect(guidance).toContain("search_files");
+    expect(guidance).toContain("read_file_content");
+  });
+
+  test("says not to browse to a vendor it has a tool for", () => {
+    // The whole point. Without this line the tool list is inert beside the browser prose.
+    expect(grantedToolGuidance(drive).toLowerCase()).toContain("do not browse");
+  });
+
+  test("says nothing at all when the Bot holds nothing", () => {
+    // A deployment with no connectors must not be told about connectors it does not have.
+    expect(grantedToolGuidance([])).toBe("");
+  });
+
+  test("a built-in Bot is told before it is told about the browser", () => {
+    const prompt = builtInAgentConfiguration(
+      {
+        id: "risk-analyst",
+        name: "Risk Analyst",
+        type: "built_in",
+        systemPrompt: "Investigate policies.",
+      },
+      { provider: "openai", defaultModel: "gpt-4.1" },
+      "openai-secret",
+      drive,
+      "BROWSER GUIDANCE HERE",
+    ).prompt as string;
+
+    // Order is the fix, not merely presence: the grants have to land before the browser prose.
+    expect(prompt.indexOf("google-drive")).toBeGreaterThan(-1);
+    expect(prompt.indexOf("google-drive")).toBeLessThan(
+      prompt.indexOf("BROWSER GUIDANCE HERE"),
+    );
+  });
+});