How connect to local installation via API or MCP via port?

[BlackFIlms]

i want use local API or MCP via port but you don't documented it.
I build comfyui to container and want connect hermes-agent from neighbor container to your tool.
How i can do it?

[g4lb]

ComfyUI has an HTTP API, not MCP. It listens only on 127.0.0.1 by default, so start it with --listen:

python main.py --listen --port 8188

Then call http://<comfyui-container>:8188/prompt from the other container. See basic_api_example.py.

[2SinceJuly]

ComfyUI has a full HTTP + WebSocket API; it just is not in the README. Everything is registered in server.py on PromptServer, so you can read the exact routes out of the source.

The endpoints that matter for an agent

Method	Path	Purpose
POST	/prompt	Queue a workflow, returns prompt_id
GET	/history/{prompt_id}	Finished output for that prompt
GET	/api/jobs/{job_id}	Job status (/api/jobs lists them)
GET	/view?filename=...&subfolder=...&type=output	Download a produced file
POST	/upload/image	Upload an input image
GET	/object_info	Node definitions + input schemas
WS	/ws?clientId=<id>	Live execution events

/object_info is the one that makes an agent usable against a workflow it did not author: it returns every node class with its input names and types, so the agent can validate widget names before submitting rather than guessing.

The flow

import json, urllib.request, uuid

client_id = str(uuid.uuid4())
payload = {"prompt": workflow_json, "client_id": client_id}

req = urllib.request.Request(
    f"http://{host}:{port}/prompt",
    data=json.dumps(payload).encode(),
    headers={"Content-Type": "application/json"},
)
prompt_id = json.load(urllib.request.urlopen(req))["prompt_id"]

# poll
with urllib.request.urlopen(f"http://{host}:{port}/history/{prompt_id}") as r:
    history = json.load(r)

client_id is optional for REST calls, but required if you want the WebSocket events to be addressed to you — the server stores it per connection and only sends executing/progress messages for that id:

self.client_id = None
...
if self.client_id == sid and self.last_node_id is not None:

Without a matching client_id on the /ws?clientId= query string, you get the broadcast frames but not the per-client progress ones.

Container-to-container networking

This is almost certainly the actual blocker, not the API. If Hermes runs in a sibling container:

1. Do not use localhost. Inside Hermes's container, 127.0.0.1 is Hermes itself. Use the ComfyUI container's service name — http://comfyui:8188 — assuming both are on the same user-defined Docker network.
2. ComfyUI must listen on 0.0.0.0. By default it binds 127.0.0.1:8188, which is unreachable from another container even with the port published. Start it with --listen 0.0.0.0 (or --listen 0.0.0.0 --port 8188).
3. They must share a network. docker network create comfy then attach both containers with --network comfy. Publishing -p 8188:8188 alone only helps the host, not a sibling container.

Quick check from inside Hermes:

curl -s http://comfyui:8188/system_stats

If that returns JSON, the network is right and any remaining problem is in the payload. If it hangs, it is a bind/network issue — try nc -vz comfyui 8188.

About MCP

ComfyUI itself does not ship an MCP server. The community MCP wrappers in the wild drive the same HTTP API above. If you want Hermes to call ComfyUI "via MCP", what you actually need is either:

• a thin MCP server you point at /prompt and /history yourself (50 lines), or
• skip MCP entirely and have the agent call the REST API — the API surface is already JSON in, JSON out.

The second is less machinery for the same result, and it removes a hop that can fail.

One gotcha

POST /prompt returns 400 with a node_errors map rather than a generic error when the workflow is malformed. Read that field — it names the exact node id and input that failed, which is what you want to feed back to an agent so it can fix and retry instead of guessing.

[BlackFIlms]

Thanks for all!
It's realy help for me!
and i think for other peoples too!
Maybe add that information to README.md?