aisthesis

aísthēsis (αἴσθησις) — Greek for "perception, sensation." The root of aesthetic.

An MCP server that gives Claude the senses it doesn't have.

Claude reasons in text. It can't make an image, speak, or listen on its own. aisthesis closes that gap by exposing a focused set of OpenAI media capabilities as Model Context Protocol tools — so any MCP client (Claude Desktop, Claude Code, and others) can generate and edit images, synthesize speech, and transcribe audio, right inside a conversation.

It is deliberately not a kitchen-sink OpenAI wrapper. Wrapping OpenAI's text models from inside Claude is redundant. The value is purely in the modalities Claude lacks natively.

Capabilities

| Tool | What it does | | --- | --- | | generate_image | Text → image (one or more), inline + optional save to disk | | edit_image | Edit / extend / inpaint an existing image, with optional mask | | text_to_speech | Text → spoken audio, selectable voice and format | | transcribe_audio | Audio file → text transcript | | transcribe_with_speakers | Transcript with speaker diarization (who spoke when) | | transcribe_with_timestamps | Time-aligned transcript: segment/word timestamps or SRT/VTT subtitles | | describe_image | Vision: caption, OCR, or answer questions about an image | | moderate_content | Pre-screen text with OpenAI's (free) moderation endpoint |

See docs/usage.md for how to use it and what the MCP server exposes, and docs/tools.md for the full per-tool reference.

Install

Requires Node ≥ 20 and your own OpenAI key (the server is bring-your-own-key — it never proxies credentials). Pick whichever path fits:

Claude Desktop Extension — one click, no terminal

Build the bundle (or download it from Releases once published):

npm install && npm run bundle   # produces aisthesis.mcpb

Double-click aisthesis.mcpb → Claude Desktop opens an installer → paste your OpenAI key in the form → done. No JSON editing, no paths.

npm / npx

Once published, point any MCP client at npx — no clone, no build:

{
  "mcpServers": {
    "aisthesis": {
      "command": "npx",
      "args": ["-y", "aisthesis"],
      "env": { "OPENAI_API_KEY": "sk-..." }
    }
  }
}

Or install it and let the wizard write your client config for you:

npm install -g aisthesis
aisthesis init      # configures Claude Desktop / Claude Code interactively

From source (development)

git clone https://github.com/igmizo/aisthesis.git
cd aisthesis && npm install && npm run build
npm run inspect     # try it in the MCP Inspector
node dist/index.js init   # or auto-configure your client

Manual config (any client): add a server that runs node /abs/path/dist/index.js with OPENAI_API_KEY in its env. Full details, including Claude Code's claude mcp add command and all env vars: docs/configuration.md.

How output works

Images and audio are always returned inline as base64 content blocks. If you set AISTHESIS_OUTPUT_DIR (or pass save_path to a tool), the result is also written to disk and the file path is returned — handy for piping into an editor or a static site.

Design notes

• BYOK, never proxied. Each user supplies their own OpenAI key, bound by their own OpenAI agreement.
• Default moderation. Content limits are enforced by OpenAI server-side; this server adds no settings to weaken them. A well-behaved server is the point.
• Configurable models. Every model is overridable via env var (see .env.example) so the defaults can age gracefully as OpenAI ships newer identifiers.

Roadmap

stdio today; a remote Cloudflare Worker variant (Streamable HTTP + R2 output) is next. See docs/roadmap.md.

License

MIT — see LICENSE.