Letz AI MCP
Official# LetzAI MCP
Two MCP servers for [LetzAI](https://letz.ai), in one repo.
| | | |
|---|---|---|
| [`server/`](server) | **Hosted** — what runs at `https://mcp.letz.ai` | Remote, Streamable HTTP, multi-user |
| [`stdio/`](stdio) | **Local** — the original Claude Desktop integration | stdio, single user, your own API key |
## Hosted server (`server/`)
The one to point an MCP client at. Add `https://mcp.letz.ai/mcp`. With OAuth enabled the client
sends you to LetzAI to log in and approve — no key to paste. A LetzAI integration token as a
bearer token works too, and always will for scripts. See
[`server/docs/authentication.md`](server/docs/authentication.md).
It is a thin, stateless adapter over the LetzAI public API: every request carries
the caller's own token, so permissions and credits resolve to that user or
organization. The server holds no key of its own. 24 tools cover image and video
generation, image editing, upscaling, trained models and user assets.
See [`server/README.md`](server/README.md) for the tool list and local development, and
[`server/docs/user-guide.md`](server/docs/user-guide.md) for the end-user setup guide.
## Local stdio server (`stdio/`)
The original integration for Claude Desktop, which runs on your own machine with
your API key in the client config. Unchanged apart from its path. Prefer the
hosted server unless you specifically want a local process.
## Deployment
The hosted server runs on GKE (`letzai-prod-services-gcp`, namespace `mcp`), behind
the `mcp.letz.ai` ingress. Kubernetes manifests live in the `infrastructure` repo
under `gcp/workloads/mcp/`.
- **Dev** — every push to `main` that touches `server/` builds and rolls out to
`mcp.dev.letz.ai` automatically.
- **Prod** — deliberate: run the *Deploy MCP (prod)* workflow, or push a `v*` tag.
TDQS
Scored across 2 tools
The two tools have completely distinct purposes: one creates images from scratch, while the other upscales existing images. There is no overlap in functionality, and an agent can easily differentiate between them based on their clear descriptions.
Both tools follow a consistent 'letzai_verb_noun' pattern with snake_case, using 'create_image' and 'upscale_image' as the core naming structure. This makes them predictable and easy to parse for an agent.
With only two tools, the server feels thin for an AI image generation domain. While create and upscale are core operations, notable gaps like editing, inpainting, or style transfer are missing, making the toolset under-scoped for typical image manipulation workflows.
The server covers basic image creation and upscaling but lacks essential operations for a complete image generation surface. There are no tools for editing, modifying, or deleting images, and advanced features like batch processing or style application are absent, leading to potential dead ends for agents.