Reddit → Muse Image → DeepSeek

Give an assistant four focused tools rather than a general-purpose shell. The assistant chooses sort, posts, image paths, and prompts; the server keeps the Reddit cookies and API key out of the tool schema.

Live services and costs

This workflow contacts Reddit and the configured OpenAI-compatible API (OpenRouter by default), downloads public media, and incurs model charges. Inspect the selected posts before using their images; respect copyright, privacy, and each service's terms. The generated image is an edit, not a byte-level patch of the original.

Set up

1. Install the MCP server, Python 3, and the Reddit CLI: cargo install reddit.
2. Download the standalone mcp.media.workflow.json into your workspace. It contains the complete inline tool implementations—no cloned repository or helper script is needed.
3. Put your browser-exported Netscape cookies.txt in the same workspace. Keep it private; do not commit it.
4. Provide your API key to the MCP server process. The tools read OPENAI_API_KEY (falling back to OPENROUTER_API_KEY when it is unset) and send requests to OPENAI_BASE_URL, which defaults to https://openrouter.ai/api/v1:

export OPENAI_API_KEY='your-key-here'
# optional: any OpenAI-compatible endpoint
export OPENAI_BASE_URL='https://openrouter.ai/api/v1'

Configure your MCP client with the media server entry from mcp.media.workflow.json. Set its working directory to the workspace containing cookies.txt (the tools create media-work/ automatically). Pass the key (and base URL) as secret environment variables, not a tool argument. If your client cannot set a working directory, start it with a wrapper that changes to your workspace before executing the server. No file other than the copied JSON and your existing cookies is required by these tool definitions; Python 3 and reddit are command-line prerequisites. The MCP binary itself remains dependency-free.

Ask the assistant

Download the first 5 new posts from the funny subreddit. Pick an image from the results and add a vivid purple border and a yellow star while keeping the scene recognizable. Then compare the original and edited images and tell me what changed.

The first tool executes the equivalent of:

reddit funny --sort new --posts 5 --cookies ./cookies.txt --out-dir media-work/reddit --no-icon

Try “top” instead of “new”, or another count from 1 to 10. The fetch_funny tool validates both fields, returns titles, URLs, and downloaded image paths, and never accepts a cookie path from the model. Posts without downloadable images have a null image path.

The generate_image tool creates an image from just a prompt. The edit_image tool submits an existing image and prompt to the OpenRouter-style Image API using input_references and the model in OPENAI_IMAGE_MODEL (default meta/muse-image). Both save the returned base64 image inside media-work/edits. The compare_images tool sends two local images and a question as image URL data payloads to the model in OPENAI_VISION_MODEL (default deepseek/deepseek-v4.1-flash) on the chat completions API. The tools return text and paths, not base64 image data to the assistant.

The edit_image and compare_images tools declare their image inputs as --fs-path-field paths with --fs-root media-work, so the server rejects any image outside media-work (including via symlinks or a leading -) before the script runs; the inline code also checks file type and size. Only use this configuration with a trusted local MCP client. It is a convenience boundary, not an OS sandbox: the child process can still read files accessible to its user. Keep the cookies and key out of other exposed workspace tools.

Run the same live test

For the repository's opt-in test, place your .env in the workspace root with OPENAI_API_KEY, OPENAI_BASE_URL=https://openrouter.ai/api/v1, and OPENAI_MODEL set to a tool-calling model. Optionally set OPENAI_IMAGE_MODEL and OPENAI_VISION_MODEL; the test passes its environment to the MCP server. In the mcp source directory run:

python3 tests/live_openrouter.py --media

The test copies only mcp.media.workflow.json and a link to cookies.txt into a temporary directory, then uses the configured LLM to select each MCP tool's arguments. It checks that five real new posts were retrieved, the edited image exists and differs from its source, the comparison model returned a substantive description, and a separate prompt generated a new image. Temporary results are removed after the test. Network and provider availability can affect this opt-in test.