> ## Documentation Index
> Fetch the complete documentation index at: https://docs.routemcp.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Stateless Chat

> Disable server-side conversation storage and supply prior turns yourself with a per-workspace stateless mode.

By default, chat is **stateful**: each `(workspace × x-session-id)` conversation is persisted server-side, and reusing the same `x-session-id` resumes it. Set `statelessMode: true` on the workspace to turn this off entirely — nothing about the conversation is stored, and the caller supplies prior turns on each request.

This is configured on the workspace endpoints (`POST /api/v1/mcp/workspaces` and `PATCH /api/v1/mcp/workspaces/{id}`). The environment (sandbox/production) is always derived from the API key — `sk_test_*` keys operate on sandbox, `sk_live_*` on production.

<Note>
  Want to route the workspace through your own LLM key and model as well? See the [Bring Your Own LLM](/guides/byo-llm) guide. The two controls are independent and can be combined on the same workspace.
</Note>

## Enable stateless mode

```bash theme={null}
curl -X PATCH https://api.routemcp.io/api/v1/mcp/workspaces/{id} \
  -H "Authorization: Bearer sk_test_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{ "statelessMode": true }'
```

`statelessMode` is a **workspace-level** boolean (default `false`). It can also be set on create. It **cannot** be overridden per chat request.

## What stops being stored

When `statelessMode` is `true`, `POST /api/v1/mcp/{workspaceId}/chat` persists **nothing**:

* no chat session is created or looked up (the `x-session-id` header is ignored),
* no messages are stored,
* no tool inputs/outputs are stored,
* no end-user context is written.

What still happens:

* **Usage metadata** (token and request counts) is still recorded for billing.
* **Analytics events are content-free** (no message text is captured).
* **`context`** is still injected into the agent for that request, but it is **not stored**.

## Supplying history

Because nothing is persisted, the caller supplies prior turns on every request via the `history` array. The **current** message stays in the normal `message` field — `history` is **prior turns only**.

```bash theme={null}
curl -X POST https://api.routemcp.io/api/v1/mcp/{workspaceId}/chat \
  -H "Authorization: Bearer sk_test_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "message": "And which one has the most stars?",
    "history": [
      { "role": "user", "message": "What repos do I have?" },
      { "role": "assistant", "message": "You have 3 repositories: api, web, and docs." }
    ]
  }'
```

| Field     | Type                         | Notes                                                                                                                                                                 |
| --------- | ---------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `history` | array of `{ role, message }` | Prior turns only (exclude the current message). `role` is `"user"` or `"assistant"`. Up to **200** items; each item's `message` is capped at \~**50,000** characters. |

<Warning>
  `history` is accepted **only** when the workspace is in `statelessMode`. Sending `history` to a non-stateless workspace returns **400** — so the caller can never silently have its supplied turns ignored.
</Warning>

<Note>
  In stateless mode there is no per-conversation end-user identity, so tool/provider connections resolve at the **workspace** level — every stateless call to the workspace shares the same connections.
</Note>

The streamed response shape is identical to stateful chat — see [the chat endpoint reference](/api-reference/endpoint/stream-chat) for the SSE event types.