diff --git a/lib/mcp-server/server.ts b/lib/mcp-server/server.ts
index d25b47e4..2129f5cb 100644
--- a/lib/mcp-server/server.ts
+++ b/lib/mcp-server/server.ts
@@ -2,6 +2,7 @@ import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { listCapabilities, ListCapabilitiesInput } from "./tools/list-capabilities";
 import { getPricing, GetPricingInput } from "./tools/get-pricing";
 import { describeCapability, DescribeCapabilityInput } from "./tools/describe-capability";
+import { runCapability, RunCapabilityInput } from "./tools/run-capability";
 import {
   capDeprecate,
   CapDeprecateInput,
@@ -440,6 +441,18 @@ export function buildStoryboardMcpServer(
     async (args: unknown) => describeCapability(args as DescribeCapabilityInput, auth),
   );
 
+  registerTool(
+    server,
+    "run_capability",
+    {
+      title: "Run any capability (output-kind generic)",
+      description:
+        "Dispatch ANY capability through the SDK — the escape hatch for caps whose output isn't media, e.g. a video-understanding / analysis cap that returns text or JSON. Media caps return a hosted URL; text/json caps return the payload inline. For ordinary image/video/audio generation prefer create_media (richer: auto-injection, quality gates). Use describe_capability first — it tells you which verb a capability wants.",
+      inputSchema: RunCapabilityInput,
+    },
+    async (args: unknown) => runCapability(args as RunCapabilityInput, auth),
+  );
+
   registerTool(
     server,
     "cap_deprecate",
diff --git a/lib/mcp-server/tools/describe-capability.ts b/lib/mcp-server/tools/describe-capability.ts
index 536af926..52b8cec9 100644
--- a/lib/mcp-server/tools/describe-capability.ts
+++ b/lib/mcp-server/tools/describe-capability.ts
@@ -33,8 +33,23 @@ const OUTPUT_TO_ACTIONS: Record<string, string[]> = {
   tool: ["generate"],
 };
 
-/** A ready-to-run create_media invocation for this cap (the stable verb). */
+// §3.5 / PR-4 — media caps invoke via create_media; everything else (text /
+// json analysis caps) via the generic run_capability verb. Used only when the
+// cap has no authored usage.invoke_via.
+const MEDIA_OUTPUT_KINDS = new Set(["image", "video", "audio", "3d"]);
+function defaultInvokeVia(outputKind: string | undefined): string {
+  return MEDIA_OUTPUT_KINDS.has(outputKind ?? "image") ? "create_media" : "run_capability";
+}
+
+/** A ready-to-run invocation for this cap via the verb its output kind wants. */
 function buildExample(entry: RegistryEntry, action: string) {
+  // Non-media (text/json analysis) caps go through the generic run_capability.
+  if (!MEDIA_OUTPUT_KINDS.has(entry.output_kind ?? "image")) {
+    return {
+      tool: "run_capability",
+      args: { capability: entry.name, prompt: "<your prompt>" } as Record<string, unknown>,
+    };
+  }
   const args: Record<string, unknown> = { action, model_override: entry.name };
   if (entry.output_kind === "video" && action === "animate") {
     args.source_url = "https://…/first-frame.png";
@@ -73,7 +88,7 @@ export async function describeCapability(input: DescribeCapabilityInput, auth: A
   // Prefer the authored usage block (cap-std §8.10) when present; else synthesize.
   const usage = entry?.usage;
   const actions = usage?.actions ?? OUTPUT_TO_ACTIONS[entry?.output_kind ?? "image"] ?? ["generate"];
-  const invokeVia = usage?.invoke_via ?? "create_media";
+  const invokeVia = usage?.invoke_via ?? defaultInvokeVia(entry?.output_kind);
   const example = usage?.example ?? (entry ? buildExample(entry, actions[0]) : undefined);
 
   const card = {
@@ -111,7 +126,9 @@ export async function describeCapability(input: DescribeCapabilityInput, auth: A
     usage?.avoid_for ? `avoid for: ${usage.avoid_for}` : null,
     entry?.display_price_usd != null ? `price: from $${entry.display_price_usd}/${entry.unit_kind ?? "unit"}` : null,
     entry?.aliases?.length ? `aliases: ${entry.aliases.join(", ")}` : null,
-    `invoke: ${invokeVia} { action: "${actions[0]}", model_override: "${name}", … }`,
+    invokeVia === "create_media"
+      ? `invoke: create_media { action: "${actions[0]}", model_override: "${name}", … }`
+      : `invoke: ${invokeVia} { capability: "${name}", … }`,
     availability === "registered" ? `⚠︎ registered but no live capacity yet — not currently routable.` : null,
   ].filter(Boolean);
 
diff --git a/lib/mcp-server/tools/run-capability.ts b/lib/mcp-server/tools/run-capability.ts
new file mode 100644
index 00000000..f734b2f8
--- /dev/null
+++ b/lib/mcp-server/tools/run-capability.ts
@@ -0,0 +1,125 @@
+/**
+ * run_capability — zero-code onboarding PR-4 (plan §4, gap G4).
+ *
+ * The output-kind-generic invocation verb. create_media is media-in/media-out;
+ * run_capability is the escape hatch for ANY capability — including text/json
+ * output (e.g. a video-understanding / screen-recording analyzer) — dispatched
+ * through the SAME transparent SDK /inference envelope, so it inherits PR-2's
+ * descriptor-driven routing (a new Live Runner app is reachable here with no
+ * client code). Media caps return a hosted URL; text/json caps return the
+ * payload inline. This is the verb describe_capability names for non-media caps.
+ *
+ * Deliberately THIN: it does not re-implement create_media's auto-injection /
+ * quality gates / project wiring. Spend-report integration is a documented
+ * follow-up (a generic escape hatch, not the primary media path).
+ */
+import { z } from "zod";
+import { extractMediaUrl, extractFalError } from "@livepeer/creative-kit";
+import { sdkPost, timeoutForCapability } from "../sdk-call";
+import type { AuthContext } from "../auth";
+import { classifyUpstreamError } from "./create-media";
+import { getRegistrySnapshot } from "@/lib/capabilities/registry-store";
+import { capabilityOutputKind } from "@/lib/sdk/capabilities";
+
+export const RunCapabilityInput = z.object({
+  capability: z
+    .string()
+    .describe("The exact capability name to run (from list_capabilities / describe_capability)."),
+  prompt: z.string().optional().describe("Text prompt / instruction, if the capability takes one."),
+  source_url: z
+    .string()
+    .optional()
+    .describe("Source media URL for image-to-X / edit / analysis capabilities."),
+  inputs: z
+    .record(z.string(), z.any())
+    .optional()
+    .describe("Extra params merged into the inference payload (capability-specific keys)."),
+  timeout: z.number().int().positive().optional().describe("Override timeout in seconds."),
+});
+export type RunCapabilityInput = z.infer<typeof RunCapabilityInput>;
+
+const MEDIA_KINDS = new Set(["image", "video", "audio", "3d"]);
+
+/** Registry-first output kind (post-PR-1/3), falling back to the static map, else text. */
+export function outputKindFor(cap: string): string {
+  const entry = getRegistrySnapshot().find((e) => e.name === cap);
+  return entry?.output_kind ?? capabilityOutputKind(cap) ?? "text";
+}
+
+type ToolResult = {
+  content: Array<{ type: "text"; text: string }>;
+  structuredContent: Record<string, unknown>;
+  isError?: true;
+};
+
+function errorResult(cap: string, raw: string): ToolResult {
+  const { hint } = classifyUpstreamError(cap, raw);
+  return {
+    isError: true,
+    content: [{ type: "text", text: `run_capability(${cap}) failed: ${raw}${hint ? ` — ${hint}` : ""}` }],
+    structuredContent: { ok: false, capability: cap, error: raw },
+  };
+}
+
+/**
+ * PURE — shape the SDK /inference result into a tool response by output kind:
+ * an upstream error → error; a media cap → hosted URL; text/json/tool → the
+ * runner payload inline. Unit-tested without a network.
+ */
+export function formatRunResult(
+  cap: string,
+  result: Record<string, unknown>,
+  outputKind: string,
+): ToolResult {
+  const topLevelError = typeof result.error === "string" ? (result.error as string) : undefined;
+  const falError = extractFalError(result);
+  if (topLevelError || falError) return errorResult(cap, topLevelError || falError!);
+
+  if (MEDIA_KINDS.has(outputKind)) {
+    const url = extractMediaUrl(result);
+    if (!url) {
+      return {
+        isError: true,
+        content: [{ type: "text", text: `run_capability(${cap}): no media URL in the response.` }],
+        structuredContent: { ok: false, capability: cap, raw: result },
+      };
+    }
+    return {
+      content: [{ type: "text", text: `${cap} → ${url}` }],
+      structuredContent: { ok: true, capability: cap, output_kind: outputKind, url },
+    };
+  }
+
+  // text / json / tool — return the runner payload inline.
+  const data = (result.data ?? result) as unknown;
+  const text = typeof data === "string" ? data : JSON.stringify(data, null, 2);
+  return {
+    content: [{ type: "text", text }],
+    structuredContent: { ok: true, capability: cap, output_kind: outputKind, result: data },
+  };
+}
+
+export async function runCapability(input: RunCapabilityInput, auth: AuthContext): Promise<ToolResult> {
+  const cap = input.capability;
+  const params: Record<string, unknown> = { ...(input.inputs ?? {}) };
+  if (input.source_url) params.source_url = input.source_url;
+
+  const timeoutMs = timeoutForCapability(cap);
+  let result: Record<string, unknown>;
+  try {
+    result = await sdkPost<Record<string, unknown>>(
+      "/inference",
+      {
+        capability: cap,
+        prompt: input.prompt ?? "",
+        params,
+        timeout: input.timeout ?? Math.ceil(timeoutMs / 1000),
+      },
+      { ...auth, timeoutMs },
+    );
+  } catch (e) {
+    return errorResult(cap, e instanceof Error ? e.message : String(e));
+  }
+
+  return formatRunResult(cap, result, outputKindFor(cap));
+}
diff --git a/tests/unit/run-capability.test.ts b/tests/unit/run-capability.test.ts
new file mode 100644
index 00000000..9d1caf26
--- /dev/null
+++ b/tests/unit/run-capability.test.ts
@@ -0,0 +1,47 @@
+import { describe, it, expect } from "vitest";
+import { formatRunResult } from "@/lib/mcp-server/tools/run-capability";
+
+describe("run_capability formatRunResult (PR-4)", () => {
+  it("media cap → hosted URL", () => {
+    const r = formatRunResult("flux-dev", { url: "https://x/a.png" }, "image");
+    expect(r.isError).toBeUndefined();
+    expect(r.structuredContent.ok).toBe(true);
+    expect(r.structuredContent.url).toBe("https://x/a.png");
+    expect(r.content[0].text).toContain("https://x/a.png");
+  });
+
+  it("media cap with NO url → error", () => {
+    const r = formatRunResult("flux-dev", { foo: 1 }, "image");
+    expect(r.isError).toBe(true);
+    expect(r.structuredContent.ok).toBe(false);
+  });
+
+  it("text cap → payload inline (from result.data)", () => {
+    const r = formatRunResult(
+      "screen-agent",
+      { data: { report_markdown: "# Bug\n- step 1", summary: "one bug" } },
+      "text",
+    );
+    expect(r.isError).toBeUndefined();
+    expect(r.structuredContent.ok).toBe(true);
+    expect(r.content[0].text).toContain("report_markdown");
+    expect((r.structuredContent.result as Record<string, unknown>).summary).toBe("one bug");
+  });
+
+  it("text cap with a bare string payload → returned verbatim", () => {
+    const r = formatRunResult("some-text-cap", { data: "hello world" }, "text");
+    expect(r.content[0].text).toBe("hello world");
+  });
+
+  it("upstream error (top-level error string) → error with the raw message", () => {
+    const r = formatRunResult("flux-dev", { error: "no orchestrator available" }, "image");
+    expect(r.isError).toBe(true);
+    expect(r.structuredContent.error).toBe("no orchestrator available");
+  });
+
+  it("json/tool output kind falls through to the inline payload path (not media)", () => {
+    const r = formatRunResult("yolo-detect", { data: { detections: [] } }, "json");
+    expect(r.structuredContent.ok).toBe(true);
+    expect(r.structuredContent.output_kind).toBe("json");
+  });
+});