Streamable HTTP in MCP: How Bidirectional Streaming Works Without SSE

Most people think MCP streaming disappeared when SSE was deprecated.

That is wrong.

Streaming did not go away — permanent connections did.

This article explains how Streamable HTTP works and why it is superior to the old SSE model.

---

The Old Model (SSE Transport)

Originally MCP used:

GET /mcp        → server → client stream (SSE)
POST /messages  → client → server commands

A permanent connection stayed open forever.
This caused:

• idle sockets
• memory pressure
• reconnection complexity
• load-balancer nightmares

---

The New Model — Streamable HTTP

Now MCP uses:

POST /mcp   ↔   streaming response

No permanent pipe.
No session IDs.
No idle sockets.

Every request owns its own stream.

---

What “Streamable” Actually Means

The server does not immediately close the POST response.

It keeps the socket open and flushes JSON-RPC messages over time.

Example:

POST /mcp
{ "method":"tool/call" }

Response:

HTTP/1.1 200 OK
Transfer-Encoding: chunked

{ "progress": 10 }
{ "progress": 50 }
{ "awaitingHuman": true, "taskId":"t123" }

The connection remains open.

---

Human-in-the-Loop Flow

1. Tool pauses

Server streams:

{ "awaitingHuman": true, "taskId":"t123" }

2. Human approves

Client opens a new POST:

POST /mcp
{ "method":"task/approve", "taskId":"t123" }

3. Server resumes the original stream

The server finds the open response stream for task t123 and continues writing:

{ "approved": true }
{ "final": "done" }

Then it closes the first POST.

---

Multiple Requests at the Same Time

Clients can run many tools concurrently:

POST /mcp   → stream A
POST /mcp   → stream B
POST /mcp   → stream C

Each is independent.

---

When Does a Stream Close?

A POST stream closes when:

• tool finishes
• error occurs
• client aborts
• server times out

No stream lives forever.

---

Why This Is Better Than SSE

SSE Transport	Streamable HTTP
One permanent pipe per client	Only pipes when work exists
Needs session map	Task-scoped streams
Hard to scale	Scales like REST
Two endpoints	One endpoint
Idle sockets everywhere	Zero idle sockets

---

The Golden Rule

Every MCP task owns its own temporary HTTP stream.
When the task ends, the connection ends.

---

Final Takeaway

Streamable HTTP gives MCP:

• real-time streaming
• human-in-loop
• concurrency
• fault tolerance

— without permanent connections.

That’s why SSE was deprecated — not because streaming was wrong, but because idle pipes were.