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

# Self-hosted secure bridge relay (enterprise)

> Own the full loop: run your own relay for the secure wss:// bridge instead of depending on a third-party service — for data residency, compliance, or a no-external-dependency requirement.

When [`connect`](./cloud-deployment) drives a remote HTTPS pod, comfyui-mcp needs a
valid-TLS `wss://` path from the pod's browser panel back to the agent bridge on
your machine (a plain `ws://` from an `https://` page is browser-blocked). By
default this is a **cloudflared quick tunnel** — zero setup, but ephemeral (a
fresh random hostname every run).

For organizations that need to **own that path end-to-end** — the control
channel never transits a third party, a fixed domain for firewall/audit rules,
or no dependency on a service outside your infrastructure — comfyui-mcp supports
routing the secure bridge through a relay **you operate**, in place of the
default quick tunnel.

<Note>
  The reference implementation is open source:
  **[artokun/comfyui-mcp-relay](https://github.com/artokun/comfyui-mcp-relay)** — a
  Cloudflare Worker + Durable Object, MIT licensed, marked as a GitHub template so
  you can click **Use this template** and have your own copy in seconds. Fork it,
  deploy it as-is, or adapt it to whatever WebSocket-capable infrastructure your
  organization already runs — the README documents the wire protocol, session
  lifecycle, and design tradeoffs (multiplexing, auth, hibernation) in full.
</Note>

## Why self-host

* **Data residency / compliance** — the bridge carries tool calls and graph
  operations (not image bytes, which flow pod↔browser directly), but some
  organizations still need that control channel to stay inside infrastructure
  they operate.
* **No third-party runtime dependency** — a fixed, owned endpoint instead of
  relying on an external relay's availability.
* **Firewall / audit rules** — a stable domain you control is easier to
  allow-list and log than an ephemeral hostname that changes every session.

## How it fits together

Both sides of the bridge dial **out** — nobody needs an inbound port open
anywhere. Your relay's job is to pair one orchestrator (running on a user's
machine) with the browser panel connections on the pod it's driving, and shuttle
bytes between them; it never needs to understand comfyui-mcp's own tool-call
protocol riding on top.

Once a relay is deployed, pointing comfyui-mcp at it is a couple of environment
variables:

```bash theme={null}
export COMFYUI_MCP_TUNNEL_BACKEND=relay
export COMFYUI_MCP_RELAY_URL=wss://your-relay.internal.example.com
# optional: a shared secret gating who can open a session on your relay at all
export COMFYUI_MCP_RELAY_KEY=your-shared-deploy-secret
```

With `COMFYUI_MCP_TUNNEL_BACKEND` unset, comfyui-mcp keeps using the default
ephemeral quick-tunnel behavior — this is entirely opt-in.

## Deploy the reference implementation

```bash theme={null}
gh repo create your-org/your-relay --template artokun/comfyui-mcp-relay --clone
cd your-relay
npm install
wrangler login
npm run deploy
```

See **[artokun/comfyui-mcp-relay](https://github.com/artokun/comfyui-mcp-relay)**
for the full README — the wire protocol, auth model, and deploy steps.