Media Gen MCP

## 🚀 media-gen-mcp

**Media Gen MCP** is a **strict TypeScript** Model Context Protocol (MCP) server designed for OpenAI Images (`gpt-image-1.5`, `gpt-image-1`), OpenAI Videos (Sora), and Google GenAI Videos (Veo). It enables users to generate and edit images, create and remix video jobs, and fetch media from URLs or disk. It offers smart `resource_link` vs inline `image` outputs and optional `sharp` processing. Focused on production, it features full strict typechecking, ESLint, and Vitest CI. It is compatible with fast-agent, Claude Desktop, ChatGPT, Cursor, VS Code, Windsurf, and any MCP-compatible client.

**Design principle**: Spec-first, type-safe image tooling that strictly adheres to the OpenAI Images API and MCP, with fully static TypeScript types and flexible result placements/response formats for different clients.

- **Generate images**: Create images from text prompts using OpenAI's `gpt-image-1.5` model (with `gpt-image-1` compatibility and DALL·E support planned for future versions).
- **Edit images**: Perform inpainting, outpainting, and compositing on 1 to 16 images at once, with advanced prompt control.
- **Generate videos**: Generate videos via OpenAI Videos (`sora-2`, `sora-2-pro`) and manage video jobs, including creation, remixing, listing, retrieval, deletion, and asset downloads.
- **Generate videos via Google GenAI**: Generate videos using Google GenAI (Veo), with operation polling and file-first downloads.
- **Fetch & compress images**: Fetch and compress images from HTTP(S) URLs or local file paths, with smart size/quality optimization.
- **Debug MCP output shapes**: Use the `test-images` tool to debug MCP output shapes, mirroring production result placement (`content`, `structuredContent`, `toplevel`).
- **Integrates with**: [fast-agent](https://github.com/strato-space/fast-agent), [Windsurf](https://windsurf.com), [Claude Desktop](https://www.anthropic.com/claude/desktop), [Cursor](https://cursor.com), [VS Code](https://code.visualstudio.com/),and any MCP-compatible client.

## ✨ Features

- **Strict MCP spec support**: The tool outputs are first-class [`CallToolResult`](https://github.com/modelcontextprotocol/spec/blob/main/schema/2025-11-25/schema.json) objects from the latest MCP schema, including `content` items (`text`, `image`, `resource_link`, `resource`), optional `structuredContent`, optional top-level `files`,and an `isError` flag for failures.
- **Full gpt-image-1 .5 and sora-2/sora-2-pro parameters coverage (generate & edit)**: 
  - [`openai-images-generate`](#openai-images-generate) mirrors the OpenAI Images [`create`](https://platform.openai.com/docs/api-reference/images/create) API for `gpt-image-1.5` (and `gpt-image-1`), covering parameters such as background handling, moderation, size, quality, output format, output compression, `n`, `user`, etc.
  - [`openai-images-edit`](#openai-images-edit) mirrors the OpenAI Images [`createEdit`](https://platform.openai.com/docs/api-reference/images/createEdit) API for `gpt-image-1.5` (and `gpt-image-1`), supporting parameters like image, mask, `n`, quality, size, `user`.
- **OpenAI Videos (Sora) job tooling (create / remix / list / retrieve / delete / content)**: 
    - [`openai-videos-create`](#openai-videos-create) mirrors [`videos/create`](https://platform.openai.com/docs/api-reference/videos/create) and can optionally wait for completion. 
    - [`openai-videos-remix`](#openai-videos-remix) mirrors [`videos/remix`](https://platform.openai.com/docs/api-reference/videos/remix). 
    - [`openai-videos-list`](#openai-videos-list) mirrors [`videos/list`](https://platform.openai.com/docs/api-reference/videos/list). 
    - [`openai-videos-retrieve`](#openai-videos-retrieve) mirrors [`videos/retrieve`](https://platform.openai.com/docs/api-reference/videos/retrieve). 
    - [`openai-videos-delete`](#openai-videos-delete) mirrors [`videos/delete`](https://platform.openai.com/docs/api-reference/videos/delete). 
    - [`openai-videos-retrieve-content`](#openai-videos-retrieve-content) mirrors [`videos/content`](https://platform.openai.com/docs/api-reference/videos/content) and downloads `video` / `thumbnail` / `spritesheet` assets to disk, returning MCP `resource_link` (default) or embedded `resource` blocks (via `tool_result`).
- **Google GenAI (Veo) operations + downloads (generate / retrieve operation / retrieve content)**: 
    - [`google-videos-generate`](#google-videos-generate) starts a long-running operation (`ai.models.generateVideos`) and can optionally wait for completion and download `.mp4` outputs. [Veo model reference](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo-video-generation). 
    - [`google-videos-retrieve-operation`](#google-videos-retrieve-operation) polls an existing operation. 
    - [`google-videos-retrieve-content`](#google-videos-retrieve-content) downloads an `.mp4` from a completed operation, returning MCP `resource_link` (default) or embedded `resource` blocks (via `tool_result`).
- **Fetch and process images from URLs or files**: The [`fetch-images`](#fetch-images) tool loads images from HTTP(S) URLs or local file paths, with optional user-controlled compression (disabled by default). It supports parallel processing of up to 20 images.
- **Fetch videos from URLs or files**: The [`fetch-videos`](#fetch-videos) tool lists local videos or downloads remote video URLs to disk and returns MCP `resource_link` (default) or embedded `resource` blocks (via `tool_result`).
- **Mix and edit up to 16 images**: The [`openai-images-edit`](#openai-images-edit) tool accepts `image` as a single string or an array of 1 - 16 file paths/base64 strings, matching the OpenAI spec for GPT Image models (`gpt-image-1.5`, `gpt-image-1`) image edits.
- **Smart image compression**: Built-in compression using [sharp](https://sharp.pixelplumbing.com/) iteratively reduces quality and dimensions to fit MCP payload limits while maintaining visual quality.
- **Resource-aware file output with `resource_link`**: 
    - Automatically switches from inline base64 to `file` when the total response size exceeds a safe threshold. 
    - Outputs are written to disk using `output_<time_t>_media-gen__<tool>_<id>.<ext>` filenames (images use a generated UUID; videos use the OpenAI `video_id`) and are exposed to MCP clients via `content[]` depending on `tool_result` (`resource_link`/`image` for images, `resource_link`/`resource` for video downloads).
- **Built-in test-images tool for MCP client debugging**: The [`test-images`](#test-images) tool reads sample images from a configured directory and returns them using the same result-building logic as production tools. Use `tool_result` and `response_format` parameters to test how different MCP clients handle `content[]` and `structuredContent`.
- **Structured MCP error handling**: All tool errors (validation, OpenAI API failures, I/O) are returned as MCP errors with `isError: true` and `content: [{ type: "text", text: <error message> }]`, making failures easy to parse and surface in MCP clients.

## 🚀 Quick Start

### Installation

```sh
git clone https://github.com/strato-space/media-gen-mcp.git
cd media-gen-mcp

npm install
npm run build

Build modes:

• npm run build: A strict TypeScript build with all strict flags enabled, including skipLibCheck: false. Incremental builds via .tsbuildinfo take approximately 2 - 3 seconds on a warm cache.
• npm run esbuild: Fast bundling via esbuild (no type checking, useful for rapid iteration).

Development mode (no build required)

For development or when TypeScript compilation fails due to memory constraints:

npm run dev  # Uses tsx to run TypeScript directly

Quality checks

npm run lint        # ESLint with typescript-eslint
npm run typecheck   # Strict tsc --noEmit
npm run test        # Unit tests (vitest)
npm run test:watch  # Watch mode for TDD
npm run ci          # lint + typecheck + test

Unit tests

The project uses vitest for unit testing. Tests are located in test/.

Covered modules:

Test categories:

• compression: isCompressionAvailable, detectImageFormat, processBufferWithCompression, readAndProcessImage
• helpers: isHttpUrl, isAbsolutePath, isBase64Image, ensureDirectoryWritable, resolveOutputPath, getResultPlacement, buildResourceLinks
• env: Config loading and validation for MEDIA_GEN_* / MEDIA_GEN_MCP_* settings
• logger: Truncation and error formatting behavior
• schemas: Validation for openai-images-*, openai-videos-*, fetch-images, fetch-videos, test-images inputs, boundary testing (prompt length, image count limits, path validation)

npm run test
# ✓ test/compression.test.ts (12 tests)
# ✓ test/helpers.test.ts (31 tests)
# ✓ test/env.test.ts (19 tests)
# ✓ test/logger.test.ts (10 tests)
# ✓ test/pricing.test.ts (5 tests)
# ✓ test/schemas.test.ts (69 tests)
# ✓ test/fetch-images.integration.test.ts (3 tests)
# ✓ test/fetch-videos.integration.test.ts (3 tests)
# Tests: 152 passed

Run directly via npx (no local clone)

You can also run the server straight from a remote repo using npx:

npx -y github:strato-space/media-gen-mcp --env-file /path/to/media-gen.env

The --env-file argument tells the server which env file to load (e.g., when you keep secrets outside the cloned directory). The file should contain OPENAI_API_KEY, optional Azure variables, and any MEDIA_GEN_MCP_* settings.

secrets.yaml (optional)

You can keep API keys (and optional Google Vertex AI settings) in a secrets.yaml file (compatible with the fast-agent secrets template):

openai:
  api_key: <your-api-key-here>
anthropic:
  api_key: <your-api-key-here>
google:
  api_key: <your-api-key-here>
  vertex_ai:
    enabled: true
    project_id: your-gcp-project-id
    location: europe-west4

media-gen-mcp loads secrets.yaml from the current working directory (or from --secrets-file /path/to/secrets.yaml) and applies it to env vars; values in secrets.yaml override env, and <your-api-key-here> placeholders are ignored.

Quick start with fast-agent and Windsurf

fast-agent

In fast-agent, MCP servers are configured in fastagent.config.yaml under the mcp.servers section (see the fast-agent docs).

To add media-gen-mcp from GitHub via npx as an MCP server:

# fastagent.config.yaml

mcp:
  servers:
    # your existing servers (e.g., fetch, filesystem, huggingface, ...)
    media-gen-mcp:
      command: "npx"
      args: ["-y", "github:strato-space/media-gen-mcp", "--env-file", "/path/to/media-gen.env"]

Put OPENAI_API_KEY and other settings into media-gen.env (see .env.sample in this repo).

Windsurf

Add an MCP server that runs media-gen-mcp from GitHub via npx using the JSON format below (similar to Claude Desktop / VS Code):

{
  "mcpServers": {
    "media-gen-mcp": {
      "command": "npx",
      "args": ["-y", "github:strato-space/media-gen-mcp", "--env-file", "/path/to/media-gen.env"]
    }
  }
}

🔑 Configuration

Add to your MCP client config (fast-agent, Windsurf, Claude Desktop, Cursor, VS Code):

{
  "mcpServers": {
    "media-gen-mcp": {
      "command": "npx",
      "args": ["-y", "github:strato-space/media-gen-mcp"],
      "env": { "OPENAI_API_KEY": "sk-..." }
    }
  }
}

Also supports Azure deployments:

{
  "mcpServers": {
    "media-gen-mcp": {
      "command": "npx",
      "args": ["-y", "github:strato-space/media-gen-mcp"],
      "env": {
        // "AZURE_OPENAI_API_KEY": "sk-...",
        // "AZURE_OPENAI_ENDPOINT": "my.endpoint.com",
        "OPENAI_API_VERSION": "2024-12-01-preview"
      }
    }
  }
}

Environment variables

• Set OPENAI_API_KEY (and optionally AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, OPENAI_API_VERSION) in the environment of the process that runs node dist/index.js (shell, systemd unit, Docker env, etc.).
• The server will optionally load a local .env file from its working directory if present (it does not override already-set environment variables).
• You can also pass --env-file /path/to/env when starting the server (including via npx); this file is loaded via dotenv before tools run, again without overriding already-set variables.

Logging and base64 truncation

To avoid flooding logs with huge image payloads, the built-in logger applies a log-only sanitizer to structured data passed to log.debug/info/warn/error:

• Truncates configured string fields (e.g., b64_json, base64, string data, image_url) to a short preview controlled by LOG_TRUNCATE_DATA_MAX (default: 64 characters). The list of keys defaults to LOG_SANITIZE_KEYS inside src/lib/logger.ts and can be overridden via MEDIA_GEN_MCP_LOG_SANITIZE_KEYS (comma-separated list of field names).
• Sanitization is applied only to log serialization; tool results returned to MCP clients are never modified.

Control via environment:

• MEDIA_GEN_MCP_LOG_SANITIZE_IMAGES (default: true):
  • 1, true, yes, on – Enable truncation (default behavior).
  • 0, false, no, off – Disable truncation and log full payloads.

Field list and limits are configured in src/lib/logger.ts via LOG_SANITIZE_KEYS and LOG_TRUNCATE_DATA_MAX.

Security and local file access

• Allowed directories: All tools are restricted to paths matching MEDIA_GEN_DIRS. If unset, defaults to /tmp/media-gen-mcp (or %TEMP%/media-gen-mcp on Windows).
• Test samples: MEDIA_GEN_MCP_TEST_SAMPLE_DIR adds a directory to the allowlist and enables the test-images tool.
• Local reads: fetch-images accepts file paths (absolute or relative). Relative paths are resolved against the first MEDIA_GEN_DIRS entry and must still match an allowed pattern.
• Remote reads: HTTP(S) fetches are filtered by MEDIA_GEN_URLS patterns. Empty = allow all.
• Writes: openai-images-generate, openai-images-edit, fetch-images, and fetch-videos write under the first entry of MEDIA_GEN_DIRS. test-images is read-only and does not create new files.

Glob patterns

Both MEDIA_GEN_DIRS and MEDIA_GEN_URLS support glob wildcards:

URL examples:

MEDIA_GEN_URLS=https://*.cdn.example.com/,https://storage.example.com/**/assets/

Path examples:

MEDIA_GEN_DIRS=/home/*/media-gen/output/,/data/**/images/

⚠️ Warning: Trailing wildcards without a delimiter (e.g., /home/user/* or https://cdn.com/**) expose entire subtrees and trigger a console warning at startup.

Recommended mitigations

1. Run under a dedicated OS user with access only to allowed directories.
2. Keep allowlists minimal. Avoid * in home directories or system paths.
3. Use explicit MEDIA_GEN_URLS prefixes for remote fetches.
4. Monitor allowed directories via OS ACLs or backups.

Tool Result Parameters: tool_result and response_format

Image tools (openai-images-*, fetch-images, test-images)

Video download tools (openai-videos-create / openai-videos-remix when downloading, openai-videos-retrieve-content, google-videos-generate when downloading, google-videos-retrieve-content, fetch-videos)

Google video tools (google-videos-*)

tool_result — controls content[]

• Images (openai-images-*, fetch-images, test-images):
  • resource_link (default): Emits ResourceLink items with file:// or https:// URIs.
  • image: Emits base64 ImageContent blocks.
• Videos (tools that download video data):
  • resource_link (default): Emits ResourceLink items with file:// or https:// URIs.
  • resource: Emits EmbeddedResource blocks with base64 resource.blob.

response_format — controls structuredContent

For OpenAI images, structuredContent always contains an OpenAI ImagesResponse-style object:

{
  "created": 1234567890,
  "data": [
    { "url": "https://..." } // or { "b64_json": "..." } depending on response_format
  ]
}

• url (default): data[].url contains file URLs.
• b64_json: data[].b64_json contains base64-encoded image data.

For Google videos, response_format controls whether structuredContent.response.generatedVideos[].video prefers:

• url (default): video.uri (and strips video.videoBytes).
• b64_json: video.videoBytes (and strips video.uri).

Backward Compatibility (MCP 5.2.6)

Per MCP spec 5.2.6, a TextContent block with serialized JSON (always using URLs in data[]) is also included in content[] for backward compatibility with clients that don't support structuredContent.

Example tool result structure:

{
  "content": [
    // ResourceLink or ImageContent based on tool_result
    { "type": "resource_link", "uri": "https://...", "name": "image.png", "mimeType": "image/png" },
    // Serialized JSON for backward compatibility (MCP 5.2.6)
    { "type": "text", "text": "{ \"created\": 1234567890, \"data\": [{ \"url\": \"https://...\" }] }" }
  ],
  "structuredContent": {
    "created": 1234567890,
    "data": [{ "url": "https://..." }]
  }
}

ChatGPT MCP client behavior (chatgpt.com, as of 2025-12-01):

• ChatGPT currently ignores content[] image data in favor of structuredContent.
• For ChatGPT, use response_format: "url" and configure the first MEDIA_GEN_MCP_URL_PREFIXES entry as a public HTTPS prefix (for example MEDIA_GEN_MCP_URL_PREFIXES=https://media-gen.example.com/media).

For Anthropic clients (Claude Desktop, etc.), the default configuration works well.

Network access via mcp-proxy (SSE)

For networked SSE access, you can front media-gen-mcp with mcp-proxy or its equivalent. This setup has been tested with the TypeScript SSE proxy implementation punkpeye/mcp-proxy.

For example, a one-line command looks like:

mcp-proxy --host=0.0.0.0 --port=99 --server=sse --sseEndpoint=/ --shell 'npx -y github:strato-space/media-gen-mcp --env-file /path/to/media-gen.env'

In production, you would typically wire this up via a systemd template unit that loads PORT/SHELL_CMD from an EnvironmentFile= (see server/mcp/mcp@.service style setups).

🛠 Tool signatures

openai-images-generate

Arguments (input schema)

• prompt (string, required): Text prompt describing the desired image. Max length: 32,000 characters.
• background ("transparent" | "opaque" | "auto", optional): Background handling mode. If background is "transparent", then output_format must be "png" or "webp".
• model ("gpt-image-1.5" | "gpt-image-1", optional, default: "gpt-image-1.5")
• moderation ("auto" | "low", optional): Content moderation behavior, passed through to the Images API.
• n (integer, optional): Number of images to generate. Min: 1, Max: 10.
• output_compression (integer, optional): Compression level (0–100). Only applied when output_format is "jpeg" or "webp".
• output_format ("png" | "jpeg" | "webp", optional): Output image format. If omitted, the server treats output as PNG semantics.
• quality ("auto" | "high" | "medium" | "low", default: "high")
• size ("1024x1024" | "1536x1024" | "1024x1536" | "auto", default: "1024x1536")
• user (string, optional): User identifier forwarded to OpenAI for monitoring.
• response_format ("url" | "b64_json", default: "url"): Response format (aligned with OpenAI Images API):
  • "url": File/URL-based output (resource_link items, image_url fields, data[].url in api placement).
  • "b64_json": Inline base64 image data (image content, data[].b64_json in api placement).
• tool_result ("resource_link" | "image", default: "resource_link"): Controls content[] shape:
  • "resource_link" emits ResourceLink items (file/URL-based).
  • "image" emits base64 ImageContent blocks.

Behavior notes

• The server uses OpenAI gpt-image-1.5 by default (set model: "gpt-image-1" for legacy behavior).
• If the total size of all base64 images would exceed the configured payload threshold (default ~50MB via MCP_MAX_CONTENT_BYTES), the server automatically switches the effective output mode to file/URL-based and saves images to the first entry of MEDIA_GEN_DIRS (default: /tmp/media-gen-mcp).
• Even when you explicitly request response_format: "b64_json", the server still writes the files to disk (for static hosting, caching, or later reuse). Exposure of file paths / URLs in the tool result then depends on MEDIA_GEN_MCP_RESULT_PLACEMENT and per-call result_placement (see section below).

Output (MCP CallToolResult, when placement includes "content")

• When the effective output mode is "base64":
  • content is an array that may contain:
    • Image items: { type: "image", data: <base64 string>, mimeType: <"image/png" | "image/jpeg" | "image/webp"> }
    • Optional text items with revised prompts returned by the Images API (for models that support it, e.g., DALL·E 3): { type: "text", text: <revised_prompt string> }
• When the effective output mode is "file":
  • content contains one resource_link item per file, plus the same optional text items with revised prompts: { type: "resource_link", uri: "file:///absolute-path-1.png", name: "absolute-path-1.png", mimeType: <image mime> }
  • For gpt-image-1.5 and gpt-image-1, an additional text line is included with a pricing estimate (based on structuredContent.usage), and structuredContent.pricing contains the full pricing breakdown.

When result_placement includes "api", openai-images-generate instead returns an OpenAI Images API-like object without MCP wrappers:

{
  "created": 1764599500,
  "data": [
    { "b64_json": "..." } // or { "url": "https://.../media/file.png" } when output: "file"
  ],
  "background": "opaque",
  "output_format": "png",
  "size": "1024x1024",
  "quality": "high"
}

openai-images-edit

Arguments (input schema)

• image (string or string[], required): Either a single absolute path to an image file (.png, .jpg, .jpeg, .webp), a base64-encoded image string (optionally as a data:image/...;base64,... URL), or an HTTP(S) URL pointing to a publicly accessible image, or an array of 1–16 such strings (for multi-image editing). When an HTTP(S) URL is provided, the server fetches the image and converts it to base64 before sending to OpenAI.
• prompt (string, required): Text description of the desired edit. Max length: 32,000 characters.
• mask (string, optional): Absolute path, base64 string, or HTTP(S) URL for a mask image (PNG < 4MB, same dimensions as the source image). Transparent areas mark regions to edit.
• model ("gpt-image-1.5" | "gpt-image-1", optional, default: "gpt-image-1.5")
• n (integer, optional): Number of images to generate. Min: 1, Max: 10.
• quality ("auto" | "high" | "medium" | "low", default: "high")
• size ("1024x1024" | "1536x1024" | "1024x1536" | "auto", default: "1024x1536")
• user (string, optional): User identifier forwarded to OpenAI for monitoring.
• response_format ("url" | "b64_json", default: "url"): Response format (aligned with OpenAI Images API):
  • "url": File/URL-based output (resource_link items, image_url fields, data[].url in api placement).
  • "b64_json": Inline base64 image data (image content, data[].b64_json in api placement).
• tool_result ("resource_link" | "image", default: "resource_link"): Controls content[] shape:
  • "resource_link" emits ResourceLink items (file/URL-based).
  • "image" emits base64 ImageContent blocks.

Behavior notes

• The server accepts image and mask as absolute paths, base64/data URLs, or HTTP(S) URLs.
• When an HTTP(S) URL is provided, the server fetches the image and converts it to a base64 data URL before calling OpenAI.
• For edits, the server always returns PNG semantics (mime type image/png) when emitting images.

Output (MCP CallToolResult)

• When the effective output mode is "base64":
  • content is an array that may contain:
    • Image items: { type: "image", data: <base64 string>, mimeType: "image/png" }
    • Optional text items with revised prompts (when the underlying model returns them): { type: "text", text: <revised_prompt string> }
• When the effective output mode is "file":
  • content contains one resource_link item per file, plus the same optional text items with revised prompts: { type: "resource_link", uri: "file:///absolute-path-1.png", name: "absolute-path-1.png", mimeType: "image/png" }
  • For gpt-image-1.5 and gpt-image-1, an additional text line is included with a pricing estimate (based on structuredContent.usage), and structuredContent.pricing contains the full pricing breakdown.

When result_placement includes "api", openai-images-edit follows the same raw API format as openai-images-generate (top-level created, data[], background, output_format, size, quality with b64_json for base64 output or url for file output).

Error handling (both tools)

• On errors inside the tool handler (validation, OpenAI API failures, I/O, etc.), the server returns a CallToolResult marked as an error:
  • isError: true
  • content: [{ type: "text", text: <error message string> }]
• The error message text is taken directly from the underlying exception message, without additional commentary from the server, while full details are logged to the server console.

openai-videos-create

Create a video generation job using the OpenAI Videos API (videos.create).

Arguments (input schema)

• prompt (string, required): Text prompt describing the video (max 32K chars).
• input_reference (string, optional): Optional image reference (HTTP(S) URL, base64/data URL, or file path).
• input_reference_fit ("match" | "cover" | "contain" | "stretch", default: "contain"): How to fit input_reference to the requested video size:
  • match: Require exact dimensions (fails fast on mismatch).
  • cover: Resize + center-crop to fill.
  • contain: Resize + pad/letterbox to fit (default).
  • stretch: Resize with distortion.
• input_reference_background ("blur" | "black" | "white" | "#RRGGBB" | "#RRGGBBAA", default: "blur"): Padding background used when input_reference_fit="contain".
• model ("sora-2" | "sora-2-pro", default: "sora-2-pro")
• seconds ("4" | "8" | "12", optional)
• size ("720x1280" | "1280x720" | "1024x1792" | "1792x1024", optional): 1024x1792 and 1792x1024 require sora-2-pro. If input_reference is omitted and size is omitted, the API default is used.
• wait_for_completion (boolean, default: true): When true, the server polls openai-videos-retrieve until completed or failed (or timeout), then downloads assets.
• timeout_ms (integer, default: 900000)
• poll_interval_ms (integer, default: 2000)
• download_variants (string[], default: ["video"]): Allowed values: "video" | "thumbnail" | "spritesheet".
• tool_result ("resource_link" | "resource", default: "resource_link"): Controls content[] shape for downloaded assets:
  • "resource_link" emits ResourceLink items (file/URL-based).
  • "resource" emits EmbeddedResource blocks with base64 resource.blob.

Output (MCP CallToolResult)

• structuredContent: OpenAI Video object (job metadata; final state when wait_for_completion=true).
• content: Includes resource_link (default) or embedded resource blocks for downloaded assets (when requested) and text blocks with JSON. Includes a summary JSON block: { "video_id": "...", "pricing": { "currency": "USD", "model": "...", "size": "...", "seconds": 4, "price": 0.1, "cost": 0.4 } | null } (and when waiting: { "video_id": "...", "assets": [...], "pricing": ... }).

openai-videos-remix

Create a remix job from an existing video_id (videos.remix).

Arguments (input schema)

• video_id (string, required)
• prompt (string, required)
• wait_for_completion, timeout_ms, poll_interval_ms, download_variants, tool_result: Same semantics as openai-videos-create (default wait is true).

openai-videos-list

List video jobs (videos.list).

Arguments (input schema)

• after (string, optional): Cursor (video id) to list after.
• limit (integer, optional)
• order ("asc" | "desc", optional)

Output

• structuredContent: OpenAI list response shape { data, has_more, last_id }.
• content: A text block with serialized JSON.

openai-videos-retrieve

Retrieve job status (videos.retrieve).

Arguments

• video_id (string, required)

openai-videos-delete

Delete a video job (videos.delete).

Arguments

• video_id (string, required)

openai-videos-retrieve-content

Retrieve an asset for a completed job (videos.downloadContent, REST GET /videos/{video_id}/content), write it under allowed MEDIA_GEN_DIRS, and return MCP resource_link (default) or embedded resource blocks (via tool_result).

Arguments (input schema)

• video_id (string, required)
• variant ("video" | "thumbnail" | "spritesheet", default: "video")
• tool_result ("resource_link" | "resource", default: "resource_link"): Controls content[] shape.

Output (MCP CallToolResult)

• structuredContent: OpenAI Video object.
• content: A resource_link (or embedded resource), a summary JSON block { video_id, variant, uri, pricing }, plus the full video JSON.

google-videos-generate

Create a Google video generation operation using the Google GenAI SDK (@google/genai) ai.models.generateVideos.

Arguments (input schema)

• prompt (string, optional)
• input_reference (string, optional): Image-to-video input (HTTP(S) URL, base64/data URL, or file path under MEDIA_GEN_DIRS).
• input_reference_mime_type (string, optional): Override for input_reference MIME type (must be image/*).
• input_video_reference (string, optional): Video-extension input (HTTP(S) URL or file path under MEDIA_GEN_DIRS; mutually exclusive with input_reference).
• model (string, default: "veo-3.1-generate-001")
• number_of_videos (integer, default: 1)
• aspect_ratio ("16:9" | "9:16", optional)
• duration_seconds (integer, optional):
  • Veo 2 models: 5–8 seconds (default: 8).
  • Veo 3 models: 4, 6, or 8 seconds (default: 8).
  • When using referenceImages: 8 seconds.
• person_generation ("DONT_ALLOW" | "ALLOW_ADULT" | "ALLOW_ALL", optional)
• wait_for_completion (boolean, default: true)
• timeout_ms (integer, default: 900000)
• poll_interval_ms (integer, default: 10000)
• download_when_done (boolean, optional; defaults to true when waiting)
• tool_result ("resource_link" | "resource", default: "resource_link"): Controls content[] shape when downloading generated videos.
• response_format ("url" | "b64_json", default: "url"): Controls structuredContent.response.generatedVideos[].video fields:
  • "url" prefers video.uri (and strips video.videoBytes).
  • "b64_json" prefers video.videoBytes (and strips video.uri).

Requirements

• Gemini Developer API: Set GEMINI_API_KEY (or GOOGLE_API_KEY), or google.api_key in secrets.yaml.
• Vertex AI: Set GOOGLE_GENAI_USE_VERTEXAI=true, GOOGLE_CLOUD_PROJECT, and GOOGLE_CLOUD_LOCATION (or google.vertex_ai.* in secrets.yaml).

Output

• structuredContent: Google operation object (includes name, done, and response.generatedVideos[] when available).
• content: Status text, optional .mp4 resource_link (default) or embedded resource blocks (when downloaded), plus JSON text blocks for compatibility.

google-videos-retrieve-operation

Retrieve/poll an existing Google video operation (ai.operations.getVideosOperation).

Arguments

• operation_name (string, required)
• response_format ("url" | "b64_json", default: "url")

Output

• structuredContent: Google operation object.
• content: JSON text blocks with a short summary + the full operation.

google-videos-retrieve-content

Download .mp4 content for a completed operation and return file-first MCP resource_link (default) or embedded resource blocks (via tool_result).

Arguments

• operation_name (string, required)
• index (integer, default: 0): Selects response.generatedVideos[index].
• tool_result ("resource_link" | "resource", default: "resource_link"): Controls content[] shape.
• response_format ("url" | "b64_json", default: "url"): Controls structuredContent.response.generatedVideos[].video fields.

Recommended workflow

1. Call google-videos-generate with wait_for_completion=true (default) to get the completed operation and downloads; set to false only if you need the operation id immediately.
2. Poll google-videos-retrieve-operation until done=true.
3. Call google-videos-retrieve-content to download an .mp4 and receive a resource_link (or embedded resource).

fetch-images

Fetch and process images from URLs or local file paths with optional compression.

Arguments (input schema)

• sources (string[], optional): Array of image sources: HTTP(S) URLs or file paths (absolute or relative to the first MEDIA_GEN_DIRS entry). Min: 1, Max: 20 images. Mutually exclusive with ids and n.
• ids (string[], optional): Array of image IDs to fetch by local filename match under the primary MEDIA_GEN_DIRS[0] directory. IDs must be safe ([A-Za-z0-9_-] only; no .., *, ?, slashes). Matches filenames containing _{id}_ or _{id}. (supports both single outputs and multi-output suffixes like _1.png). When ids is used, compression and file are not supported (no new files are created). Mutually exclusive with sources and n.
• n (integer, optional): When set, returns the last N image files from the primary MEDIA_GEN_DIRS[0] directory. Files are sorted by modification time (most recently modified first). Mutually exclusive with sources and ids.
• compression (object, optional):
  • max_size (integer, optional): Max dimension in pixels. Images larger than this will be resized.
  • max_bytes (integer, optional): Target max file size in bytes. Default: 819200 (800KB).
  • quality (integer, optional): JPEG/WebP quality 1 - 100. Default: 85.
  • format ("jpeg" | "png" | "webp", optional): Output format. Default: jpeg.
• response_format ("url" | "b64_json", default: "url"): Response format: file/URL-based (url) or inline base64 (b64_json).
• tool_result ("resource_link" | "image", default: "resource_link"): Controls content[] shape:
  • "resource_link" emits ResourceLink items (file/URL-based).
  • "image" emits base64 ImageContent blocks.
• file (string, optional): Base path for output files. If multiple images, index suffix is added.

Behavior notes

• Images are processed in parallel for maximum throughput.
• Compression is only applied when compression options are provided.
• Compression uses sharp with iterative quality/size reduction when enabled.
• Partial success: If some sources fail, successful images are still returned with errors listed in the response.
• When n is provided, it is only honored when the MEDIA_GEN_MCP_ALLOW_FETCH_LAST_N_IMAGES environment variable is set to true. Otherwise, the call fails with a validation error.
• Sometimes an MCP client (for example, ChargeGPT) may not wait for a response from media-gen-mcp due to a timeout. In creative environments where you need to quickly retrieve the latest openai-images-generate / openai-images-edit outputs, you can use fetch-images with the n argument. When the MEDIA_GEN_MCP_ALLOW_FETCH_LAST_N_IMAGES=true environment variable is set, fetch-images will return the last N files from MEDIA_GEN_DIRS[0] even if the original generation or edit operation timed out on the MCP client side.

fetch-videos

Fetch videos from HTTP(S) URLs or local file paths.

Arguments (input schema)

• sources (string[], optional): Array of video sources: HTTP(S) URLs or file paths (absolute or relative to the first MEDIA_GEN_DIRS entry). Min: 1, Max: 20 videos. Mutually exclusive with ids and n.
• ids (string[], optional): Array of video IDs to fetch by local filename match under the primary MEDIA_GEN_DIRS[0] directory. IDs must be safe ([A-Za-z0-9_-] only; no .., *, ?, slashes). Matches filenames containing _{id}_ or _{id}. (supports both single outputs and multi-asset suffixes like _thumbnail.webp). When ids is used, file is not supported (no downloads; returns existing files). Mutually exclusive with sources and n.
• n (integer, optional): When set, returns the last N video files from the primary MEDIA_GEN_DIRS[0] directory. Files are sorted by modification time (most recently modified first). Mutually exclusive with sources and ids.
• tool_result ("resource_link" | "resource", default: "resource_link"): Controls content[] shape:
  • "resource_link" emits ResourceLink items (file/URL-based).
  • "resource" emits EmbeddedResource blocks with base64 resource.blob.
• file (string, optional): Base path for output files (used when downloading from URLs). If multiple videos are downloaded, an index suffix is added.

Output

• content: One resource_link (default) or embedded resource block per resolved video, plus an optional error summary text block.
• structuredContent: { data: [{ source, uri, file, mimeType, name, downloaded }], errors?: string[] }.

Behavior notes

• URL downloads are only allowed when the URL matches MEDIA_GEN_URLS (when set).
• When n is provided, it is only honored when the MEDIA_GEN_MCP_ALLOW_FETCH_LAST_N_VIDEOS environment variable is set to true. Otherwise, the call fails with a validation error.

test-images

Debug tool for testing MCP result placement without calling OpenAI API.

Enabled only when MEDIA_GEN_MCP_TEST_SAMPLE_DIR is set. The tool reads existing images from this directory and does not create new files.

Arguments (input schema)

• response_format ("url" | "b64_json", default: "url")
• result_placement ("content" | "api" | "structured" | "toplevel" or array of these, optional): Override MEDIA_GEN_MCP_RESULT_PLACEMENT for this call.
• compression (object, optional): Same logical tuning knobs as fetch-images, but using camelCase keys:
  • maxSize (integer, optional): Max dimension in pixels.
  • maxBytes (integer, optional): Target max file size in bytes.
  • quality (integer, optional): JPEG/WebP quality 1–100.
  • format ("jpeg" | "png" | "webp", optional): Output format.
• tool_result ("resource_link" | "image", default: "resource_link"): Controls content[] shape:
  • "resource_link" emits ResourceLink items (file/URL-based).
  • "image" emits base64 ImageContent blocks.

Behavior notes

• Reads up to 10 images from the sample directory (no sorting — filesystem order).
• Uses the same result-building logic as openai-images-generate and openai-images-edit (including result_placement overrides).
• When output == "base64" and compression is provided, sample files are read and compressed in memory using sharp; original files on disk are never modified.
• Useful for testing how different MCP clients handle various result structures.
• When result_placement includes "api", the tool returns a mock OpenAI Images API-style object:
  • Top level: created, data[], background, output_format, size, quality.
  • For response_format: "b64_json" each data[i] contains b64_json.
  • For response_format: "url" each data[i] contains url instead of b64_json.

Debug CLI helpers for test-images

For local debugging, there are two helper scripts that call test-images directly:

• npm run test-images: Uses debug/debug-call.ts and prints the validated CallToolResult as seen by the MCP SDK client. Usage:

npm run test-images -- [placement] [--response_format url|b64_json]
# examples:
# npm run test-images -- structured --response_format b64_json
# npm run test-images -- structured --response_format url

• npm run test-images:raw: Uses debug/debug-call-raw.ts and prints the raw JSON-RPC result (the underlying CallToolResult without extra wrapping). Same CLI flags as above.

Both scripts truncate large fields for readability:

• image_url → first 80 characters, then ...(N chars);
• b64_json and data (when it is a base64 string) → first 25 characters, then ...(N chars).

🧩 Version policy

Semantic Versioning (SemVer)

This package follows SemVer: MAJOR.MINOR.PATCH (x.y.z).

• MAJOR: Breaking changes (tool names, input schemas, output shapes).
• MINOR: New tools or backward-compatible additions (new optional params, new fields in responses).
• PATCH: Bug fixes and internal refactors with no intentional behavior change.

Since 1.0.0, this project follows standard SemVer rules: breaking changes bump MAJOR (npm’s ^1.0.0 allows 1.x, but not 2.0.0).

Dependency policy

This repository aims to stay closely aligned with current stable releases:

• MCP SDK: Targeting the latest stable @modelcontextprotocol/sdk and schema.
• OpenAI SDK: Regularly updated to the latest stable openai package.
• Zod: Using the Zod 4.x line (currently ^4.1.3). In this project, we previously ran on Zod 3.x and, in combination with the MCP TypeScript SDK typings, hit heavy TypeScript errors when passing .shape into inputSchema — in particular TS2589 ("type instantiation is excessively deep and possibly infinite") and TS2322 (schema shape not assignable to AnySchema | ZodRawShapeCompat). We track the upstream discussion in modelcontextprotocol/typescript-sdk#494 and the related Zod typing work in colinhacks/zod#5222, and keep the stack on a combination that passes full strict compilation reliably.
• Tooling stack (Node.js, TypeScript, etc.): Developed and tested against recent LTS / current releases, with a dedicated tsconfig-strict.json that enables all strict TypeScript checks (strict, noUnusedLocals, noUnusedParameters, exactOptionalPropertyTypes, noUncheckedIndexedAccess, noPropertyAccessFromIndexSignature, etc.).

You are welcome to pin or downgrade Node.js, TypeScript, the OpenAI SDK, Zod, or other pieces of the stack if your environment requires it, but please keep in mind:

• We primarily test and tune against the latest stack.
• Issues that only reproduce on older runtimes / SDK versions may be harder for us to investigate and support.
• Upstream compatibility is validated first of all against the latest MCP spec and OpenAI Images API.

This project is intentionally a bit futuristic: it tries to keep up with new capabilities as they appear in MCP and OpenAI tooling (in particular, robust multimodal/image support over MCP and in ChatGPT’s UI). A detailed real‑world bug report and analysis of MCP image rendering in ChatGPT is listed in the References section as a case study.

If you need a long-term-stable stack, pin exact versions in your own fork and validate them carefully in your environment.

🧩 Typed tool callbacks

All tool handlers use strongly typed callback parameters derived from Zod schemas via z.input<typeof schema>:

// Schema definition
const openaiImagesGenerateBaseSchema = z.object({
  prompt: z.string().max(32000),
  background: z.enum(["transparent", "opaque", "auto"]).optional(),
  // ... more fields
});

// Type alias
type OpenAIImagesGenerateArgs = z.input<typeof openaiImagesGenerateBaseSchema>;

// Strictly typed callback
server.registerTool(
  "openai-images-generate",
  { inputSchema: openaiImagesGenerateBaseSchema.shape, ... },
  async (args: OpenAIImagesGenerateArgs, _extra: unknown) => {
    const validated = openaiImagesGenerateSchema.parse(args);
    // ... handler logic
  },
);

This pattern provides:

• Static type safety: IDE autocomplete and compile-time checks for all input fields.
• Runtime validation: Zod .parse() ensures all inputs match the schema before processing.
• MCP SDK compatibility: inputSchema: schema.shape provides the JSON Schema for tool registration.

All tools (openai-images-*, openai-videos-*, fetch-images, fetch-videos, test-images) follow this pattern.

🧩 Tool annotations

This MCP server exposes the following tools with annotation hints:

These hints help MCP clients understand that these tools:

• May invoke external APIs or read external resources (open world).
• Do not modify existing project files or user data; they only create new media files (images/videos) in configured output directories.
• May produce different outputs on each call, even with the same inputs.

Because readOnlyHint is set to true for most tools, MCP platforms (including chatgpt.com) can treat this server as logically read-only and usually will not show "this tool can modify your files" warnings.

📁 Project structure

media-gen-mcp/
├── src/
│   ├── index.ts              # MCP server entry point
│   └── lib/
│       ├── compression.ts    # Image compression (sharp)
│       ├── env.ts            # Env parsing + allowlists (+ glob support)
│       ├── helpers.ts        # URL/path validation, result building
│       ├── logger.ts         # Structured logging + truncation helpers
│       └── schemas.ts        # Zod schemas for all tools
├── test/
│   ├── compression.test.ts             # 12 tests
│   ├── env.test.ts                     # 19 tests
│   ├── fetch-images.integration.test.ts# 2 tests
│   ├── fetch-videos.integration.test.ts# 2 tests
│   ├── helpers.test.ts                 # 31 tests
│   ├── logger.test.ts                  # 10 tests
│   └── schemas.test.ts                 # 64 tests
├── debug/                    # Local debug helpers (MCP client scripts)
├── plan/                     # Design notes / plans
├── dist/                     # Compiled output
├── tsconfig.json
├── vitest.config.ts
├── package.json
├── CHANGELOG.md
├── README.md
└── AGENTS.md

📝 License

MIT

🩺 Troubleshooting

• Make sure your OPENAI_API_KEY is valid and has image API access.
• You must have a verified OpenAI organization. After verifying, it can take 15–20 minutes for image API access to activate.
• File paths [optional param] must be absolute:
  • Unix/macOS/Linux: Starting with / (e.g., /path/to/image.png).
  • Windows: Drive letter followed by : (e.g., C:/path/to/image.png or C:\path\to\image.png).
• For file output, ensure the target directory is writable.
• If you see errors about file types, check your image file extensions and formats.

🙏 Inspiration

This server was originally inspired by SureScaleAI/openai-gpt-image-mcp, but is now a separate implementation focused on closely tracking the official specifications:

• OpenAI Images API alignment: The arguments for openai-images-generate and openai-images-edit mirror images.create / gpt-image-1.5: prompt, n, size, quality, background, output_format, output_compression, user, plus response_format (url / b64_json) with the same semantics as the OpenAI Images API.
• MCP Tool Result alignment (image + resource_link): With result_placement = "content", the server follows the MCP 5.2 Tool Result section (5.2.2 Image Content, 5.2.4 Resource Links) and emits strongly-typed content[] items:
  • { "type": "image", "data": "<base64>", "mimeType": "image/png" } for response_format = "b64_json".
  • { "type": "resource_link", "uri": "file:///..." | "https://...", "name": "...", "mimeType": "image/..." } for file/URL-based output.
• Raw OpenAI-style API output: With result_placement = "api", the tool result itself is an OpenAI Images-style object: { created, data: [...], background, output_format, size, quality, usage? }, where each data[] entry contains either b64_json (for response_format = "b64_json") or url (for response_format = "url"). No MCP wrapper fields (content, structuredContent, files, urls) are added in this mode.

In short, this library:

• Tracks the OpenAI Images API for arguments and result shape when result_placement = "api" with response_format = "url" | "b64_json".
• Follows the MCP specification for tool result content blocks (image, resource_link, text) when result_placement = "content".

Recommended presets for common clients

Default mode / Claude Desktop / strict MCP clients

For clients that strictly follow the MCP spec, the recommended (and natural) configuration is:

• result_placement = content
• response_format = b64_json

In this mode, the server returns:

• content[] with type: "image" (base64 image data) and type: "resource_link" (file/URL links), matching MCP section 5.2 (Image Content and Resource Links). This output works well for direct integration with Claude Desktop and any client that fully implements the 2025‑11‑25 spec.

chatgpt.com Developer Mode

For running this server as an MCP backend behind ChatGPT Developer Mode, the most practical configuration is the one that most closely matches the OpenAI Images API:

• result_placement = api
• response_format = url

In this mode, the tool result matches the images.create / gpt-image-1.5 format (including data[].url), which simplifies consumption from backends and libraries that expect the OpenAI schema.

However, even with this OpenAI-native shape, the chatgpt.com client does not currently render images. This behavior is documented in detail in the following report: https://github.com/strato-space/report/issues/1

⚠️ Limitations & Large File Handling

• Configurable payload safeguard: By default, this server uses a ~50MB budget (52,428,800 bytes) for inline content to stay within typical MCP client limits. You can override this threshold by setting the MCP_MAX_CONTENT_BYTES environment variable to a higher (or lower) value.
• Auto-Switch to File Output: If the total image base64 size exceeds the configured threshold, the tool automatically saves images to disk and returns file path(s) via resource_link instead of inline base64. This helps avoid client-side "payload too large" errors while still delivering full-resolution images.
• Default File Location: If you do not specify a file path, outputs are saved under MEDIA_GEN_DIRS[0] (default: /tmp/media-gen-mcp) using names like output_<time_t>_media-gen__<tool>_<id>.<ext>.
• Environment Variables:
  • MEDIA_GEN_DIRS: Set this to control where outputs are saved. Example: export MEDIA_GEN_DIRS=/your/desired/dir. This directory may coincide with your public static directory if you serve files directly from it.
  • MEDIA_GEN_MCP_URL_PREFIXES: Optional comma-separated HTTPS prefixes for public URLs, matched positionally to MEDIA_GEN_DIRS entries. When set, the server builds public URLs as <prefix>/<relative_path_inside_root> and returns them alongside file paths (for example via resource_link URIs and structuredContent.data[].url when response_format: "url"). Example: export MEDIA_GEN_MCP_URL_PREFIXES=https://media-gen.example.com/media,https://media-gen.example.com/samples
  • Best Practice: For large or production images, always use file output and ensure your client is configured to handle file paths. Configure MEDIA_GEN_DIRS and (optionally) MEDIA_GEN_MCP_URL_PREFIXES to serve images via a public web server (e.g., nginx).

🌐 Serving generated files over HTTPS

If you want ChatGPT (or any MCP client) to mention publicly accessible URLs alongside file paths:

1. Expose your image directory via HTTPS. For example, on nginx:

server {
    # listen 443 ssl http2;
    # server_name <server_name>;

    # ssl_certificate     <path>;
    # ssl_certificate_key <path>;

    location /media/ {
        alias /home/username/media-gen-mcp/media/;
        autoindex off;
        expires 7d;
        add_header Cache-Control "public, immutable";
    }
}

2. Ensure the first entry in MEDIA_GEN_DIRS points to the same directory (e.g., MEDIA_GEN_DIRS=/home/username/media-gen-mcp/media/ or MEDIA_GEN_DIRS=media/ when running from the project root).
3. Set MEDIA_GEN_MCP_URL_PREFIXES=https://media-gen.example.com/media so the server returns matching HTTPS URLs in top-level urls, resource_link URIs, and image_url fields (for response_format: "url").

Both openai-images-generate and openai-images-edit now attach files + urls for base64 and file response modes, allowing clients to reference either the local filesystem path or the public HTTPS link. This is particularly useful while ChatGPT cannot yet render MCP image blocks inline.

📚 References

• Model Context Protocol
  • MCP Specification
  • MCP Schema (2025-11-25)
• OpenAI Images
  • Images API overview
  • Images generate (gpt-image-1.5)
  • Images edit (createEdit)
  • Tools guide: image generation & revised_prompt
• OpenAI Videos
  • Videos API overview
• Case studies
  • MCP image rendering in ChatGPT (GitHub issue)
    • Symptoms: ChatGPT often ignored or mishandled MCP