The log tells us two separate things are still wrong:

| Symptom | Root cause | 1‑line fix |
|---------|------------|------------|
| `RuntimeWarning: coroutine 'batch_mcp_flow' was never awaited` | `batch_mcp_flow()` is **async**, but you call it like a plain function. | `out = await batch_mcp_flow(...)` *or* wrap the call in `asyncio.run(…)` if you want to stay sync. |
| `[MCP‑error] Expecting value: line 1 column 1 (char 0)` | The Fast‑MCP mock server replied, but your client tried to `json.loads()` an **empty body** (usually means the HTTP endpoint or headers are wrong). | Point the URL at the server **root** (`http://0.0.0.0:8000`) and send plain JSON‑RPC. |

Below is a minimal patch that removes both errors without touching the rest of your code‑base.

---

## 1 ️⃣ — make `mcp_execution_flow()` await the coroutine

```python
# ── swarms/structs/agent.py ─────────────────────────────────────────

import asyncio          # add this at the top if it’s not there

# …

def mcp_execution_flow(self, payload: dict):
    """
    Forward the tool‑call dict to every MCP server in self.mcp_servers.
    Returns the first non‑error response as a string.
    """
    try:
        # batch_mcp_flow is async ➜ run it synchronously
        result = asyncio.run(
            batch_mcp_flow(self.mcp_servers, payload)
        )
        return any_to_str(result)
    except Exception as err:
        logger.error(f"MCP flow failed: {err}")
        return f"[MCP-error] {err}"
```

*(If you prefer your whole agent to stay async, just make
`mcp_execution_flow` itself `async def` and `await batch_mcp_flow` —
then call it with `await` from the `_run` loop.)*

---

## 2 ️⃣ — use the correct Fast‑MCP endpoint

In the client that spins up the **math agent**:

```python
math_server = MCPServerSseParams(
    url="http://0.0.0.0:8000",          #  ←  root!  no “/mcp”
    headers={"Content-Type": "application/json"},
    timeout=5.0,
    sse_read_timeout=30.0,
)
```

⚠️ Fast‑MCP’s `run(transport="sse", port=8000)` already exposes both
the SSE stream and the JSON‑RPC POST endpoint on that root URL.  
Adding `/mcp` makes the request hit a 404, so the body is empty — that’s
exactly what the JSON decoder complained about.

---

## 3 ️⃣ — (optional) turn streaming off until everything works

```python
math_agent = Agent(
    # …
    streaming_on=False   #   ← easier to debug; turn back on later
)
```

With streaming disabled, `LiteLLM` returns plain strings, so your
`parse_llm_output()` method won’t be handed a
`CustomStreamWrapper` object any more.

---

### Quick test matrix

| Action | Expected result |
|--------|-----------------|
| `curl -X POST http://0.0.0.0:8000 -d '{"tool_name":"add","a":2,"b":3}' -H 'Content-Type: application/json'` | `{"result":5}` |
| Run `mock_math_server.py` | “Starting Mock Math Server on port 8000…” |
| Run `mcp_client.py`, type `add 2 and 3` | Agent replies something like “2 + 3 = 5”. No coroutine warning. |

As soon as the round‑trip works once, you can:

* re‑enable `streaming_on=True` and teach `parse_llm_output()` to turn a
  `CustomStreamWrapper` into text (`"".join(token.choices[0].delta.content for token in wrapper)`);
* point the agent at your real MCP servers instead of the mock one.

Happy hacking!