AI Hub Dev Template

This repo provides a template to kickstart development with AI Hub.

Contents

• ./skills - agent skills with information on using AI hub for AI agents. Move these to a suitable location for your preferred AI coding agent.
• ./src/Sample - Basic sample classes for tools, toolsets, agents and MCP servers. These are installed with zpm when the container is build.
• ./src/Python - An example stdio MCP server defined in Python and used in the IRIS Toolsets

Using the template

Download AI Hub Container

1. Download an AI Hub container from the Early Access Program Portal. The docker-containers end with docker.tar.gz, ensure you choose the version suitable for your operating system (arm64 for macOS).

OR

1. Copy AI Hub Container from your Flash Drive

2. Load the image with:

   docker load -i /path/to/iris-community-2026.2.0AI.162.0-docker.tar.gz

   Once it's complete you should see Loaded image: docker.iscinternal.com/docker-intersystems/intersystems/iris-community:2026.2.0AI.162.0 (if not you can use docker images to find the image name).

3. Change the Image name in the Dockerfile to match your version and operating system (image name printed above).

Build Template Repo

4. Clone this repo:

git clone https://github.com/intersystems-community/ai-hub-dev-template
cd ai-hub-dev-template

5. Add an OPENAI_API_KEY to a file called .env in this repo. You can see an example in .env.example. If you want to use another provider, change the Sample.Agent class in src/. If you don't want to use any agents at the moment (e.g. you want to create MCP tools to access from outside the container), create an empty .env file (touch .env) or remove the env_file tag from docker-compose.yml

6. Build the container with:

docker-compose up -d --build 

Using IRIS AI Hub Container

Accessing IRIS

You can find the Management Portal at http://localhost:52773/csp/sys/UtilHome.csp.

Login with: - SuperUser / SYS

You can access the IRIS Terminal with:

docker-compose exec -it iris iris session iris

or the bash terminal with:

docker-compose exec -it iris iris session iris

Testing Sample agent

There is a basic agent in src/Sample.Agent, a simple way to use it from objectscript is to run the following (note this does require an OPENAI_API_KEY to be added to .env before running th container).

zn "IRISAPP"
Set agent = ##class(Sample.Agent).%New()
Set sc = agent.%Init()
write:sc'=1 $SYSTEM.Status.GetErrorText(sc), !

Set session = agent.CreateSession()

Set request = "What tools do you have?"
Set response = agent.Chat(session, request)
Do ##class(%AI.System).RenderMarkdown(response.Content)

You can also use the agents in streaming mode as follows:

// Create Stream Renderer
Set renderer = ##class(%AI.Shell.StreamRenderer).%New()

// Request requires using both tools defined in Sample.Tools and packaged in Sample.ToolSet
Set request = "Add a person named Peter aged 16, and then get people younger than 35."

// Stream Response
Set response = agent.StreamChat(session, request , renderer, "OnChunk")

Try an interactive Chat Shell

do ##class(%AI.System).Shell("openai", $System.Util.GetEnviron("OPENAI_API_KEY"), "gpt-5-nano", "Sample.ToolSet")

Test MCP Server

The build process installs an MCP server web application at http://localhost:52773/mcp/sample. You can check this MCP server is running by going to http://localhost:52773/mcp/sample/v1/services.

For the MCP Server to be usable, there is an additional step of starting this via a Rust binary which connects to IRIS through the web gateway protocol. The Binary is installed in /usr/irissys/bin (should already be in PATH).

A sample configuration is shown in config.toml, which serves a remote HTTP server on port 8080 (which is exposed by the docker-compose file). Please note, the port for the remote HTTP server is not the same as the web server port!

To start the transport, open a bash terminal within the container:

docker-compose exec -it iris bash 

Then start the iris-mcp-server

iris-mcp-server -c config.toml run 

You can now connect the MCP server to your MCP Client of choice (e.g. coding agents like claude code) using the address: http://localhost:8080/mcp/sample.

An example python MCP client is shown in test_mcp.py, which uses Langchain's MCP adapters module. To try this, run:

pip install langchain-mcp-adapters
python test_mcp.py