image-gen-mcp
Configurable, provider-agnostic image generation, as a DeepSeek Harness
plugin. Point it at any OpenAI-compatible /v1/images/generations endpoint
and the model gains image_generate and image_config_status — plus a bundled
skill that tells it when to use them.
A Codex plugin that exposes image generation tools through MCP with a
configurable base URL and API key.
Install
DeepSeek Harness Desktop — open Plugins in the sidebar, choose Add
plugin, and enter:
https://github.com/bauerelizabeth07139/image-gen-mcp
Then switch the new dsh-image-gen-mcp bundle on, and give it a backend (see
Configuration — without a base URL and key the plugin
deliberately mounts nothing and says so).
dsh CLI — install it into the profile you actually boot:
dsh plugin --profile web add bauerelizabeth07139/image-gen-mcp
No git on the machine? pnpm resolves a git shorthand with git ls-remote,
which fails with 'git' is not recognized when git is missing. Use the tarball
instead — that path is plain HTTPS:
dsh plugin --profile web add https://codeload.github.com/bauerelizabeth07139/image-gen-mcp/tar.gz/master
The same address works in the Desktop Add plugin dialog. Replace master
with a commit SHA to pin an exact revision (/tar.gz/<sha>).
Uninstall with dsh plugin --profile web remove dsh-image-gen-mcp.
Requirements
- Node.js ≥ 22 — the harness itself; the server is plain Node with no npm
dependencies, so there is nothing to install.
- A backend: a base URL that answers
POST {baseUrl}/v1/images/generations
and an API key for it. The base URL must not include /v1.
Tools
| Tool | What it does |
|---|
mcp__image_generation__image_generate | Generates an image from a prompt. Parameters: prompt (required), model, size, n. |
mcp__image_generation__image_config_status | Reports whether the base URL and key are configured — ask this first when a generation fails for configuration reasons. |
Bundled skill
skills/image-generation/SKILL.md is registered by the plugin, so the model
knows the workflow: check image_config_status, generate with
image_generate, return the artifact or an actionable error. It is adapted from
the repository's original Codex skill — same workflow, harness tool names, plus
a note about base64 payloads.
Configuration
| Key | Environment variable | Default | Meaning |
|---|
baseUrl | IMAGE_GEN_BASE_URL | (none — required) | backend base URL, without /v1 |
apiKey | IMAGE_GEN_API_KEY | (none — required) | credential for that backend |
model | IMAGE_GEN_DEFAULT_MODEL | (empty) | default model; the server picks dall-e-3 in openai mode and gpt-image-1 in generic mode when this is empty |
provider | IMAGE_GEN_PROVIDER | generic | generic (URL response) or openai (base64 response) |
timeoutMs | IMAGE_GEN_TIMEOUT_MS | 30000 | the server's own HTTP budget |
toolCallTimeoutMs | — | 120000 | DSH's per-call budget |
env | — | {} | raw environment passthrough |
Example loader row:
- id: dsh-image-gen-mcp
name: 'dsh-image-gen-mcp'
config:
baseUrl: 'https://api.example.com'
apiKey: 'sk-...'
model: 'gpt-image-1'
provider: 'generic'
The credential is forwarded explicitly: the harness scrubs credential-shaped
variables (KEY, TOKEN, SECRET, PASSWORD) out of the environment a child
would inherit, so an exported IMAGE_GEN_API_KEY is read at load time and
written into the child's environment by the plugin.
Why this plugin ships a bridge
The server frames its stdio transport the LSP way — Content-Length: <n>
followed by a blank line — while the MCP client inside DeepSeek Harness reads
newline-delimited JSON. Mounted directly, the handshake would simply time out.
bridge.mjs translates in both directions: NDJSON in from the harness,
Content-Length frames to the server, and the reverse on the way back. Message
bodies are reframed byte for byte, never re-parsed.
How it is mounted
index.js registers the skill, then mounts scripts/server.mjs through the
bridge over stdio with failOnStartupError: true. If baseUrl or apiKey is
missing it logs a warning and mounts nothing — a server that exits at startup
would otherwise fail the whole bundle load — while still registering the skill,
so the model can tell the user exactly what to configure.
Development
npm test
Four suites, all Harness-free: the manifest and card metadata
(test/plugin.test.mjs), the bridge's framing and round trip
(test/bridge.test.mjs), the stdio mount and credential forwarding
(test/mount.test.mjs), and the bundled skill's registration
(test/skill.test.mjs). The repository's own Python validators still work:
python scripts/validate_plugin.py ., python scripts/test_server.py,
python tests/test_provider_modes.py.
Repository layout
| Path | Purpose |
|---|
index.js | the DSH plugin: skill provider + stdio mount |
bridge.mjs | NDJSON ↔ Content-Length translation |
cordis.patch.yml | the loader row that activates the plugin |
skills/image-generation/ | the bundled skill, adapted from the original |
locale/{en,zh}.json, assets/icon.svg | card title, description and artwork |
scripts/server.mjs | the MCP server, unchanged |
scripts/server.py, scripts/test_server.py | the original Python HTTP debug server and its test, unchanged |
scripts/validate_plugin.py, tests/ | the original Codex-plugin validators and tests, unchanged |
.mcp.json, .codex-plugin/ | the original Codex plugin manifest, unchanged |
README.opencode.md | the repository's original README, verbatim |
Other hosts (unchanged)
{
"mcpServers": {
"image-generation": {
"command": "node",
"args": ["scripts/server.mjs"],
"env": {
"IMAGE_GEN_BASE_URL": "${IMAGE_GEN_BASE_URL}",
"IMAGE_GEN_API_KEY": "${IMAGE_GEN_API_KEY}"
}
}
}
}
Two caveats carried over from the original setup: the relative
scripts/server.mjs needs the repository root as the working directory, and a
host that frames with newline-delimited JSON (like DSH) needs the bridge:
node bridge.mjs -- node scripts/server.mjs.
License
MIT — the repository's own licence file, unchanged.
