diff --git a/.agents/plugins/marketplace.json b/.agents/plugins/marketplace.json
new file mode 100644
index 0000000..fb49922
--- /dev/null
+++ b/.agents/plugins/marketplace.json
@@ -0,0 +1,20 @@
+{
+  "name": "comfyui-agents",
+  "interface": {
+    "displayName": "ComfyUI Agents"
+  },
+  "plugins": [
+    {
+      "name": "comfyui-agents",
+      "source": {
+        "source": "local",
+        "path": "."
+      },
+      "policy": {
+        "installation": "AVAILABLE",
+        "authentication": "ON_USE"
+      },
+      "category": "Developer Tools"
+    }
+  ]
+}
diff --git a/.codex-plugin/mcp.json b/.codex-plugin/mcp.json
new file mode 100644
index 0000000..b18f814
--- /dev/null
+++ b/.codex-plugin/mcp.json
@@ -0,0 +1,8 @@
+{
+  "mcpServers": {
+    "comfy": {
+      "type": "http",
+      "url": "http://127.0.0.1:9188/mcp"
+    }
+  }
+}
diff --git a/.codex-plugin/plugin.json b/.codex-plugin/plugin.json
new file mode 100644
index 0000000..5e61db9
--- /dev/null
+++ b/.codex-plugin/plugin.json
@@ -0,0 +1,37 @@
+{
+  "name": "comfyui-agents",
+  "version": "0.1.0",
+  "description": "Bridge an external coding agent (Codex) to a running ComfyUI graph: exchange text and images on named channels and trigger workflow runs over MCP.",
+  "author": {
+    "name": "Ethan Fel",
+    "email": "ethan.fel@ts-pc.fr"
+  },
+  "homepage": "https://github.com/ethanfel/ComfyUI-Nodes-Agents",
+  "repository": "https://github.com/ethanfel/ComfyUI-Nodes-Agents",
+  "license": "MIT",
+  "keywords": [
+    "comfyui",
+    "mcp",
+    "agent",
+    "image-generation",
+    "bridge"
+  ],
+  "skills": "./skills/",
+  "mcpServers": "./.codex-plugin/mcp.json",
+  "interface": {
+    "displayName": "ComfyUI Agent Bridge",
+    "shortDescription": "Exchange text/images with a running ComfyUI graph and trigger runs over MCP",
+    "longDescription": "Connects Codex to the ComfyUI-Nodes-Agents in-process MCP server (HTTP, 127.0.0.1:9188/mcp). Pull images/text the graph emitted, push text/images back into the graph, list channels, and run saved workflows from the agent.",
+    "developerName": "Ethan Fel",
+    "category": "Developer Tools",
+    "capabilities": [
+      "Read",
+      "Write"
+    ],
+    "defaultPrompt": [
+      "Pull the latest image from the ComfyUI graph and describe it.",
+      "Push a prompt to ComfyUI and run the txt2img workflow."
+    ],
+    "brandColor": "#1F6FEB"
+  }
+}
diff --git a/skills/comfyui-bridge/SKILL.md b/skills/comfyui-bridge/SKILL.md
new file mode 100644
index 0000000..7888e8e
--- /dev/null
+++ b/skills/comfyui-bridge/SKILL.md
@@ -0,0 +1,54 @@
+---
+name: comfyui-bridge
+description: Use when exchanging text or images with a running ComfyUI graph or triggering a ComfyUI workflow run from the agent. Trigger terms - ComfyUI, comfy_pull, comfy_push, comfy_run_workflow, named channel, inbox, outbox, render a prompt, edit the graph's image.
+license: MIT
+---
+
+# ComfyUI Agent Bridge
+
+Connects you to a running ComfyUI graph through the `comfy` MCP server
+(HTTP at `http://127.0.0.1:9188/mcp`). You exchange text and images with the
+graph on **named channels** and can trigger saved workflow runs.
+
+ComfyUI must be running with the `ComfyUI-Nodes-Agents` custom node installed,
+or the tools below will fail to connect.
+
+## Named-channel model
+
+Each channel name (default `main`) has two directions:
+
+- **inbox** = graph -> agent: an `Agent Emit` node writes the latest text/image;
+  you read it with `comfy_pull`.
+- **outbox** = agent -> graph: you write with `comfy_push`; an `Agent Receive`
+  node reads it. Each pushed value is **consumed once** (turn-based).
+
+Use distinct channel names for distinct streams (e.g. `in`, `out`, `prompt`,
+`render`); two `Agent Receive` nodes on one channel compete for the same push.
+
+## The five tools
+
+- `comfy_pull(channel="main")` -> `{turn, text, image_path}`: read what the graph
+  emitted on a channel.
+- `comfy_push(channel="main", text=None, image_path=None)`: send text/image to
+  the graph's `Agent Receive` node on a channel.
+- `comfy_list_channels()`: list active channels and their turn counters.
+- `comfy_run_workflow(name, inputs=None, wait=True)`: load `workflows/<name>.json`
+  (API-format export), optionally inject `inputs` keyed `"<node_id>.<field>"`
+  (e.g. `{"6.text": "a red fox"}`), queue it, and (with `wait=True`) poll until
+  outputs appear.
+- `comfy_get_result(prompt_id)`: fetch produced images/outputs for a queued
+  prompt id.
+
+Images cross by file path: `comfy_pull` returns an `image_path` you can read;
+pass a local PNG path to `comfy_push(image_path=...)`.
+
+## The core loop
+
+1. `comfy_push(channel, text=..., image_path=...)` — stage inputs for the graph.
+2. `comfy_run_workflow(name, wait=True)` — trigger the run and wait for outputs.
+3. `comfy_pull(channel)` — read the text/image the graph emitted back.
+
+Full render round-trip: with a workflow whose `Agent Receive(channel=prompt)`
+feeds the prompt and whose output goes to `Agent Emit(channel=render)`:
+`comfy_push("prompt", "a red fox")` -> `comfy_run_workflow("txt2img", wait=True)`
+-> `comfy_pull("render")`.