Algorithm servers¶

Algorithm servers run the heavy models — segmentation, detection, restoration — behind a single uniform interface. Each wraps one model and ships as a Docker container, usually run on a machine with a GPU. Your agent calls them through biopb; you generally just need one running and reachable.

Available pre-built servers¶

biopb publishes ready-to-run images:

Image	Model	Use
`jiyuuchc/cellpose`	Cellpose Cyto3	General cell/nucleus segmentation
`jiyuuchc/cellpose-sam`	Cellpose-SAM	Segmentation with the SAM backbone
`jiyuuchc/ucell`	ucell	Cell segmentation
`jiyuuchc/unifmir`	UNiFMIR	Universal fluorescence image restoration

Requirements¶

An NVIDIA GPU with a recent driver (>= 525).
The NVIDIA Container Toolkit, so Docker containers can use the GPU.

Running a server¶

# Standard: expose the gRPC port 50051
docker run --gpus=all -p 50051:50051 jiyuuchc/cellpose

# Use a different host port
docker run --gpus=all -p 8888:50051 jiyuuchc/cellpose

# Require a token to access
docker run --gpus=all -p 50051:50051 jiyuuchc/cellpose --token

# Debug mode, no token
docker run --gpus=all -p 50051:50051 jiyuuchc/cellpose --no-token --debug

Once it's running, update your biopb-mcp config file, i.e., ~/.config/biopb-mcp/config.json, and add its URL to the mcp.services.process_image_servers list:

{
  "mcp": {
    "services": {
      "process_image_servers": ["grpc://your-server:50051"]
    }
  }
}

so your agent knows to use it. You could of course also ask your agent to change the config file for you instead of editing by hand.

Transport is unencrypted (HTTP) by default

Run algorithm servers on a trusted network. To use TLS, put a reverse proxy (e.g. Nginx) in front to forward the gRPC calls.

Common Gotha

If dataset is large, biopb will not send all the pixels to the algorithm server but just a link to the data server - but that means you need to configure your data server to listen for the connection. All are fine if you run everything on the same machine. But if you are ever setting up a distributed deployment, pay attention to which ports you are binding!

Building your own¶

Algorithm servers are deliberately uniform: each is a thin model wrapper on a shared base, so adding a new one is mostly "wrap a model and point it at the protocol." If your lab has a custom model, see the biopb-server build docs.