diff --git a/pyproject.toml b/pyproject.toml
index b0eee33539..23ee12fd1d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -51,7 +51,7 @@ dependencies = [
     "tomli-w>=1.0.0",
     "renderers[multimodal]>=0.1.13.dev17",
     "typing_extensions>=4.12.2",
-    "mcp==2.0.0",
+    "mcp==2.2.0",
 ]
 
 [dependency-groups]
diff --git a/tests/integration/fixtures/tool_error.py b/tests/integration/fixtures/tool_error.py
new file mode 100644
index 0000000000..a18a1b08d3
--- /dev/null
+++ b/tests/integration/fixtures/tool_error.py
@@ -0,0 +1,64 @@
+"""tool-error (v1, MCP tool): a tool's exception text reaches the model.
+
+Task tools raise to tell the model what went wrong (an unknown id, an invalid argument),
+so the exception's message must come back as the tool result. The tool always raises
+with a token the prompt never reveals, so the reward is 1.0 only if the error text
+reached the trace.
+"""
+
+import verifiers.v1 as vf
+from verifiers.v1.types import content_text
+
+ERROR_TOKEN = "err-5c2"  # only the raised message carries this
+
+
+class OrdersToolset(vf.Toolset[vf.ToolsetConfig]):
+    TOOL_PREFIX = "orders"
+
+    @vf.tool
+    def lookup(self, order_id: str) -> str:
+        """Look up an order by its id."""
+        raise ValueError(f"Order {order_id} not found [{ERROR_TOKEN}]")
+
+
+class ToolErrorTaskConfig(vf.TaskConfig):
+    tools: vf.ToolsetConfig = vf.ToolsetConfig()
+
+
+class ToolErrorTask(vf.Task[vf.TaskData, vf.State, ToolErrorTaskConfig]):
+    @classmethod
+    def toolsets(cls, config: ToolErrorTaskConfig) -> list[vf.Toolset]:
+        return [OrdersToolset(config.tools)]
+
+    @vf.reward(weight=1.0)
+    async def error_reported(self, trace: vf.Trace) -> float:
+        results = (content_text(m.content) for m in trace.tool_messages)
+        return float(any(ERROR_TOKEN in r for r in results))
+
+
+class ToolErrorConfig(vf.TasksetConfig):
+    task: ToolErrorTaskConfig = ToolErrorTaskConfig()
+
+
+class ToolErrorTaskset(vf.Taskset[ToolErrorTask, ToolErrorConfig]):
+    def load(self) -> list[ToolErrorTask]:
+        return [
+            ToolErrorTask(
+                vf.TaskData(
+                    idx=0,
+                    prompt=(
+                        "Call the `lookup` tool from the `orders` MCP server with the order "
+                        'id "A-17", then reply with exactly what it returns inside '
+                        "<answer></answer> tags."
+                    ),
+                ),
+                self.config.task,
+            )
+        ]
+
+
+__all__ = ["ToolErrorTaskset"]
+
+
+if __name__ == "__main__":
+    OrdersToolset.run()
diff --git a/tests/integration/test_multi_turn_tool.py b/tests/integration/test_multi_turn_tool.py
index b3a61c7504..7c02a416f3 100644
--- a/tests/integration/test_multi_turn_tool.py
+++ b/tests/integration/test_multi_turn_tool.py
@@ -2,7 +2,8 @@
 never reveals (echo-tool), so reward 1.0 proves the tool ran. Every MCP harness reaches
 it colocated across the harness x runtime grid; the chat loop reaches it across its
 placements (colocated in the harness's box, in a Compose service beside it, or its own
-runtime) on the runtime x runtime grid."""
+runtime) on the runtime x runtime grid. A tool that raises (tool-error) reports its message
+to the model; that is server-side, so the chat loop covers it."""
 
 import pytest
 from grid import assert_one_branch, grid, placements
@@ -64,6 +65,16 @@ async def test_tool_placement(run, harness_runtime, tool_runtime):
     assert "message" in echo_tool.parameters.get("properties", {})
 
 
+async def test_tool_error(run):
+    """A tool's exception message comes back to the model, not just that the call failed."""
+    (episode,) = await run("tool-error", max_turns=4)
+    (trace,) = episode.traces
+    assert trace.ok, [error.message for error in trace.errors]
+    assert trace.reward == 1.0, [
+        message.model_dump_json(exclude_none=True) for message in trace.tool_messages
+    ]
+
+
 @pytest.mark.parametrize(
     "placement",
     [
diff --git a/uv.lock b/uv.lock
index ef901111b2..838ccf2d60 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1,5 +1,5 @@
 version = 1
-revision = 4
+revision = 5
 requires-python = ">=3.11, <3.15"
 resolution-markers = [
     "python_full_version >= '3.14' and sys_platform == 'win32'",
@@ -21,15 +21,15 @@ exclude-newer = "0001-01-01T00:00:00Z" # This has no effect and is included for
 exclude-newer-span = "P7D"
 
 [options.exclude-newer-package]
-prime-pydantic-config = false
+certifi = false
 harbor = "2026-10-08T00:00:00Z"
-ty = "2026-07-28T00:00:00Z"
-ruff = "2026-07-28T00:00:00Z"
+prime-pydantic-config = false
 prime-runs = false
-prime-tunnel = false
-certifi = false
 prime-sandboxes = false
+prime-tunnel = false
 renderers = false
+ruff = "2026-07-28T00:00:00Z"
+ty = "2026-07-28T00:00:00Z"
 
 [[package]]
 name = "aiofile"
@@ -2278,7 +2278,7 @@ wheels = [
 
 [[package]]
 name = "mcp"
-version = "2.0.0"
+version = "2.2.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "anyio" },
@@ -2296,22 +2296,22 @@ dependencies = [
     { name = "typing-inspection" },
     { name = "uvicorn", marker = "sys_platform != 'emscripten'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/74/33/32d4dff2c95bb5d897c3ef4c83649a08996b17b58f0a326d2495d4c81179/mcp-2.0.0.tar.gz", hash = "sha256:0f440e735c13ece8bb19bc62cf0b86f4313448432fbb77d35e14034f4e050728", size = 1662284, upload-time = "2026-07-28T13:45:32.346Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/76/31/ac54fb0fdd5b37de704486e288bba4fbbb463f24cfcfedbede407b854513/mcp-2.2.0.tar.gz", hash = "sha256:2dc37ecb1974becdcebdbf7561e7c15a07dbbf20ba21ba16c3593b3038b3afbd", size = 4084129, upload-time = "2026-09-07T16:06:23.439Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/67/72/7d7897418912c1d12e87556630dfb7bf0eac71160e9bef8b447960804ee3/mcp-2.0.0-py3-none-any.whl", hash = "sha256:1cb4c75d2d2c7b8c1d756355e5d82a39f2822cc7f13e22a2051d7ca3592349d6", size = 349980, upload-time = "2026-07-28T13:45:28.853Z" },
+    { url = "https://files.pythonhosted.org/packages/1b/ff/8e7eade68b8a28f7da0ed1085544341b51f9c935dbf6b95c76b7edfea6a0/mcp-2.2.0-py3-none-any.whl", hash = "sha256:bde982589473a060ae145e3406e9a5333fe538c97229ba841f5a7f92be004f81", size = 365656, upload-time = "2026-09-07T16:06:19.711Z" },
 ]
 
 [[package]]
 name = "mcp-types"
-version = "2.0.0"
+version = "2.2.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "pydantic" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/bb/56/9b8e1c152f61f6c6b07c4b5896c88c7d0ae90bac6ee6306f852fcc5c1eb0/mcp_types-2.0.0.tar.gz", hash = "sha256:d7d939b9285c9961ae8866ba75ef85da34d12bafe276efbf4eb6a131786d8379", size = 66632, upload-time = "2026-07-28T13:45:33.804Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/ae/91/762d7755d971aff8a28d75f7961656148edf27875c8026e6385aaab08ae7/mcp_types-2.2.0.tar.gz", hash = "sha256:d3ed53703ddd10d9c6399f29d322bb66f3f67ab41348ac8556ba23e07fedefad", size = 65892, upload-time = "2026-09-07T16:06:25.187Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/f5/4c/c78d78c3d52b0ac594ad7cc8ef5972adfe070e3597a8a4c6ce0cd39196ea/mcp_types-2.0.0-py3-none-any.whl", hash = "sha256:6b2de797ca2797f568b79529e1b25948e34de511bcc0bd82fef1039a6d1b8eb0", size = 69649, upload-time = "2026-07-28T13:45:30.713Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/d7/6ffba5d8cd5dd9b8a19478875c50e04945314ba5074e84d749283f27f62d/mcp_types-2.2.0-py3-none-any.whl", hash = "sha256:ea476b73ee86709ab5abc9452385ed36cc05907e582355622e294595c9a04f13", size = 69106, upload-time = "2026-09-07T16:06:21.461Z" },
 ]
 
 [[package]]
@@ -5173,6 +5173,7 @@ wheels = [
 [[package]]
 name = "verifiers"
 source = { editable = "." }
+default-groups = ["dev", "examples"]
 dependencies = [
     { name = "aiohttp" },
     { name = "anthropic" },
diff --git a/verifiers/v1/mcp/server.py b/verifiers/v1/mcp/server.py
index 2ea101e338..85c0e54588 100644
--- a/verifiers/v1/mcp/server.py
+++ b/verifiers/v1/mcp/server.py
@@ -233,9 +233,19 @@ async def wrapper(*args, **kwargs):
             token = _call_state.set(state)
             try:
                 before = self._state_adapter.dump_json(state) if sync_state else None
-                result = fn(*args, **kwargs)
-                if inspect.isawaitable(result):
-                    result = await result
+                try:
+                    result = fn(*args, **kwargs)
+                    if inspect.isawaitable(result):
+                        result = await result
+                except Exception as e:
+                    from mcp.server.mcpserver.exceptions import ToolError
+                    from mcp.shared.exceptions import MCPError
+
+                    if isinstance(e, (ToolError, MCPError)):
+                        raise
+                    # MCPServer withholds an unexpected exception's text from the client, but
+                    # tools raise to tell the model what went wrong.
+                    raise ToolError(str(e)) from e
                 if before is not None:
                     await self._push_state(before)
                 return result