diff --git a/packages/mcp-server/src/tools.test.ts b/packages/mcp-server/src/tools.test.ts
index 01773dbe33b3..2e03b5914680 100644
--- a/packages/mcp-server/src/tools.test.ts
+++ b/packages/mcp-server/src/tools.test.ts
@@ -546,6 +546,126 @@ describe("paperclip MCP tools", () => {
     expect(JSON.parse(String(init.body))).toEqual({});
   });
 
+  it("routes approval withdraw to the requester-scoped withdraw endpoint", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(
+      mockJsonResponse({ id: "approval-1", status: "withdrawn" }),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    const tool = getTool("paperclipApprovalDecision");
+    await tool.execute({
+      approvalId: "55555555-5555-5555-5555-555555555555",
+      action: "withdraw",
+      reason: "Superseded by PR #1190.",
+    });
+
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    const [url, init] = fetchMock.mock.calls[0] as [string, RequestInit];
+    expect(String(url)).toBe(
+      "http://localhost:3100/api/approvals/55555555-5555-5555-5555-555555555555/withdraw",
+    );
+    expect(init.method).toBe("POST");
+    expect(JSON.parse(String(init.body))).toEqual({ reason: "Superseded by PR #1190." });
+  });
+
+  it("falls back to decisionNote as the withdraw reason", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(
+      mockJsonResponse({ id: "approval-1", status: "withdrawn" }),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    const tool = getTool("paperclipApprovalDecision");
+    await tool.execute({
+      approvalId: "55555555-5555-5555-5555-555555555555",
+      action: "withdraw",
+      decisionNote: "Question answered itself.",
+    });
+
+    const [, init] = fetchMock.mock.calls[0] as [string, RequestInit];
+    expect(JSON.parse(String(init.body))).toEqual({ reason: "Question answered itself." });
+  });
+
+  it("prefers reason over decisionNote when a withdraw supplies both", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(
+      mockJsonResponse({ id: "approval-1", status: "withdrawn" }),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    const tool = getTool("paperclipApprovalDecision");
+    await tool.execute({
+      approvalId: "55555555-5555-5555-5555-555555555555",
+      action: "withdraw",
+      reason: "Superseded by PR #1190.",
+      decisionNote: "stale note from an earlier draft",
+    });
+
+    // `reason` is the withdraw-specific field, so it wins; `decisionNote` is only
+    // a fallback. Pinning this keeps the precedence from silently inverting.
+    const [, init] = fetchMock.mock.calls[0] as [string, RequestInit];
+    expect(JSON.parse(String(init.body))).toEqual({ reason: "Superseded by PR #1190." });
+  });
+
+  it("folds reason into decisionNote on non-withdraw actions so it is never dropped", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(mockJsonResponse({ id: "approval-1" }));
+    vi.stubGlobal("fetch", fetchMock);
+
+    const tool = getTool("paperclipApprovalDecision");
+    await tool.execute({
+      approvalId: "55555555-5555-5555-5555-555555555555",
+      action: "approve",
+      reason: "Looks good.",
+    });
+
+    // `reason` is advertised on the shared schema, so a board caller can reach for
+    // it on any action. Without the fold-back the note is silently elided by
+    // JSON.stringify and the server receives {}.
+    const [url, init] = fetchMock.mock.calls[0] as [string, RequestInit];
+    expect(String(url)).toBe(
+      "http://localhost:3100/api/approvals/55555555-5555-5555-5555-555555555555/approve",
+    );
+    expect(JSON.parse(String(init.body))).toEqual({ decisionNote: "Looks good." });
+  });
+
+  it("refuses a withdraw with a blank reason instead of dropping it", async () => {    const fetchMock = vi.fn();
+    vi.stubGlobal("fetch", fetchMock);
+
+    const tool = getTool("paperclipApprovalDecision");
+    const response = await tool.execute({
+      approvalId: "55555555-5555-5555-5555-555555555555",
+      action: "withdraw",
+      reason: "   ",
+    });
+
+    // Fails loudly and locally: the audit trail relies on the reason to tell a
+    // moot request from an abandoned one, so this must never reach the server
+    // as a withdrawal with no note.
+    expect(response.isError).toBe(true);
+    expect(response.content[0]?.text).toContain("withdraw requires a non-empty reason");
+    expect(fetchMock).not.toHaveBeenCalled();
+  });
+
+  it("leaves requester and pending scoping to the server for withdraw", async () => {
+    // The route rejects withdrawing another agent's card (403) and the service
+    // rejects withdrawing an already-decided one (409). Pin that the tool sends
+    // no actor or status override that could widen either check.
+    const fetchMock = vi.fn().mockResolvedValue(
+      mockJsonResponse({ error: "Only requesting agent can withdraw this approval" }, 403),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    const tool = getTool("paperclipApprovalDecision");
+    const response = await tool.execute({
+      approvalId: "55555555-5555-5555-5555-555555555555",
+      action: "withdraw",
+      reason: "Not mine to withdraw.",
+    });
+
+    const [, init] = fetchMock.mock.calls[0] as [string, RequestInit];
+    expect(Object.keys(JSON.parse(String(init.body)))).toEqual(["reason"]);
+    expect(response.isError).toBe(true);
+    expect(response.content[0]?.text).toContain("Only requesting agent can withdraw this approval");
+  });
+
   it("rejects invalid generic request paths", async () => {
     vi.stubGlobal("fetch", vi.fn());
 
diff --git a/packages/mcp-server/src/tools.ts b/packages/mcp-server/src/tools.ts
index 5b532d630bdb..c14f398945f3 100644
--- a/packages/mcp-server/src/tools.ts
+++ b/packages/mcp-server/src/tools.ts
@@ -194,8 +194,15 @@ const createRequestCheckboxConfirmationToolSchema = z.object({
 
 const approvalDecisionSchema = z.object({
   approvalId: approvalIdSchema,
-  action: z.enum(["approve", "reject", "requestRevision", "resubmit"]),
+  action: z.enum(["approve", "reject", "requestRevision", "resubmit", "withdraw"]),
   decisionNote: z.string().optional(),
+  // `withdraw` and `resubmit` are both requester-scoped; only approve/reject/
+  // requestRevision call assertBoard. The withdraw route additionally requires a
+  // non-empty reason, so accept a dedicated `reason` rather than making callers
+  // learn that `decisionNote` is overloaded. `decisionNote` is still read as a
+  // fallback, and on the non-withdraw path `reason` folds back into
+  // `decisionNote` so a note can never be silently dropped either way.
+  reason: z.string().optional(),
   payloadJson: z.string().optional(),
 });
 
@@ -729,9 +736,26 @@ export function createToolDefinitions(client: PaperclipApiClient): ToolDefinitio
     ),
     makeTool(
       "paperclipApprovalDecision",
-      "Approve, reject, request revision, or resubmit an approval",
+      "Approve, reject, request revision, resubmit, or withdraw an approval. `approve`, `reject`, and `requestRevision` are board-only — an agent calling them gets `403 Board access required`. `withdraw` and `resubmit` are **both requester-scoped**: the requesting agent may rescind its own ask or resubmit it, so a card that went moot is yours to clear rather than something to ask a human to close, and a card the board sent back as `revision_requested` is yours to resubmit. You can only act on cards you filed, and only while they are still pending; acting on another agent's card is refused (403), as is acting on one already decided (409). `withdraw` requires a non-empty `reason` (or `decisionNote`) — the audit trail relies on it to tell a moot request apart from an abandoned one. Note one destructive side effect: withdrawing a `hire_agent` approval also terminates the pending agent it would have created (it would otherwise be stranded frozen with no approval left to decide it).",
       approvalDecisionSchema,
-      async ({ approvalId, action, decisionNote, payloadJson }) => {
+      async ({ approvalId, action, decisionNote, reason, payloadJson }) => {
+        if (action === "withdraw") {
+          // Refuse here rather than letting an empty reason reach the server as a
+          // bare 400: the caller learns which field to fill, and a withdrawal can
+          // never silently lose the note the audit trail depends on.
+          const withdrawReason = (reason ?? decisionNote ?? "").trim();
+          if (!withdrawReason) {
+            throw new Error(
+              "withdraw requires a non-empty reason: pass `reason` (or `decisionNote`) saying why the request became moot",
+            );
+          }
+          return client.requestJson(
+            "POST",
+            `/approvals/${encodeURIComponent(approvalId)}/withdraw`,
+            { body: { reason: withdrawReason } },
+          );
+        }
+
         const path =
           action === "approve"
             ? `/approvals/${encodeURIComponent(approvalId)}/approve`
@@ -747,7 +771,7 @@ export function createToolDefinitions(client: PaperclipApiClient): ToolDefinitio
             ? replacementPayload === undefined
               ? {}
               : { payload: replacementPayload }
-            : { decisionNote };
+            : { decisionNote: decisionNote ?? reason };
 
         return client.requestJson("POST", path, { body });
       },
diff --git a/skills/paperclip/SKILL.md b/skills/paperclip/SKILL.md
index 6683f02334e4..5eab47a347e7 100644
--- a/skills/paperclip/SKILL.md
+++ b/skills/paperclip/SKILL.md
@@ -251,6 +251,23 @@ POST /api/companies/{companyId}/approvals
 
 `issueIds` links the approval into the issue thread. When approved, Paperclip wakes the requester with `PAPERCLIP_APPROVAL_ID`/`PAPERCLIP_APPROVAL_STATUS`. Keep the payload concise and decision-ready.
 
+### Withdrawing an approval you filed
+
+If a card you filed goes moot — the work landed another way, the question answered itself, the ask was wrong — **withdraw it yourself**. Do not comment asking the board to close it: a pending approval sits in a human's queue until someone acts on it, and the retraction comment costs them a read on top of the card.
+
+```json
+POST /api/approvals/{approvalId}/withdraw
+{ "reason": "Superseded by PR #1190, which landed the same patch on 08-09." }
+```
+
+Or via MCP: `paperclipApprovalDecision` with `action: "withdraw"` and a `reason`.
+
+Scope: **the requesting agent, on its own still-pending card.** Withdrawing another agent's card is refused (403); withdrawing one the board already decided is refused (409). `reason` is required and must be non-empty — the audit trail uses it to tell a moot request apart from an abandoned one.
+
+One destructive side effect to know before you reach for this: withdrawing a `hire_agent` approval **also terminates the pending agent it would have created**. That is deliberate — the agent is parked in `pending_approval` and would otherwise be stranded frozen with no approval left to decide it — but it is not obvious from the word "withdraw". Requester-scoping means you can only ever terminate a hire you filed yourself.
+
+`resubmit` is scoped the same way: if the board sends your card back as `revision_requested`, **you resubmit it yourself** — that is not a board-only action either. Only `approve`, `reject`, and `requestRevision` are board-only and return `403 Board access required` for agents. That 403 is about those three actions, not about approvals generally — it does not mean you cannot retract or resubmit your own ask.
+
 ## Issue-Thread Interactions
 
 Issue-thread interactions are first-class cards that render in the issue thread and capture a typed board/user response. Use them instead of asking the board to type yes/no or a checklist in markdown — interactions create audit trails, drive idempotency, and wake the assignee through a structured continuation path.