The Model Context Protocol is a new specification and multi-language implementation for allowing LLMs to integrate and interact with other services.
Tools are an important part of the specification, as they provide AI/LLMs with the ability to take actions, providing standardized meta-data (aka tool descriptions) that the AI/LLM can use to decide upon, provide input for and then take actions based upon the model context.
The creation of MCP servers has/is growing very quickly. Currently, however, most of these implementations do not support the ability to add and remove tools from MCP servers dynamically, and require a new server (and a new set of implemented tools) to used by clients.
Enter Remote Tools for MCP Servers. This repo makes available a Remote Services distribution provider based upon Py4j and Python <-> Java RPC framework used by Apache Spark and others. Combined with iPOPO, which is a component framework for Python, and includes RSA and a py4j distribution provider, it's now easy to use Remote Services to dynamically extend/enhance MCP python servers (FastMCP below) with new Java implemented tools.
MCP Client (example: inspector) <-- MCP --> MCP Server (example: Python) <-- ArithmeticTools remote service --> Remote ArithmeticTools Server (example: Java)
Here is an example Java interface providing an ArithmeticTools declaration with MCP annotations:
public interface ArithmeticTools extends ToolGroupService {
@Tool(description = "computes the sum of the two integer arguments")
@ToolAnnotations(destructiveHint = true, title="howdy")
@ToolResult(description = "the integer result for this tool")
int add(@ToolParam(description = "a is the first argument") int a, @ToolParam(description = "b is the second argument") int b);
@Tool(description = "return the product of the two given integer arguments named a and b")
int multiply(@ToolParam(description = "a is the first argument") int a, @ToolParam(description = " b is the second argument") int b);
}
Here is a simple implementation of this interface. Notice that neither the service interface nor implementation class refer to any classes that are part of RS/RSA, the distribution provider, karaf, or any other framework.
In Karaf 4.4.6 this example and all supporting code can be installed with these two karaf console commands
karaf@root()> feature:repo-add https://download.eclipse.org/rt/ecf/latest/site.p2/karaf-features.xmlthen
karaf@root()> feature:install ecf-rs-examples-python.java-mcp-toolserviceThis will produce debug output to your karaf console
14:39:05.818;EXPORT_REGISTRATION;exportedSR=[org.eclipse.ecf.examples.ai.mcp.toolservice.api.ArithmeticTools];cID=URIID [uri=py4j://127.0.0.1:25333/java];rsId=1
--Endpoint Description---
<endpoint-descriptions xmlns="http://www.osgi.org/xmlns/rsa/v1.0.0">
<endpoint-description>
<property name="ecf.endpoint.id" value-type="String" value="py4j://127.0.0.1:25333/java"/>
<property name="ecf.endpoint.id.ns" value-type="String" value="ecf.namespace.py4j"/>
<property name="ecf.endpoint.ts" value-type="Long" value="1748727545801"/>
<property name="ecf.rsvc.id" value-type="Long" value="1"/>
<property name="endpoint.framework.uuid" value-type="String" value="82798710-8368-461b-b07d-0250e9e064cc"/>
<property name="endpoint.id" value-type="String" value="74b58995-4e59-4936-8743-4599066c1b10"/>
<property name="endpoint.service.id" value-type="Long" value="146"/>
<property name="objectClass" value-type="String">
<array>
<value>org.eclipse.ecf.examples.ai.mcp.toolservice.api.ArithmeticTools</value>
</array>
</property>
<property name="osgi.ds.satisfying.condition.target" value-type="String" value="(osgi.condition.id=true)"/>
<property name="remote.configs.supported" value-type="String">
<array>
<value>ecf.py4j.host</value>
</array>
</property>
<property name="remote.intents.supported" value-type="String">
<array>
<value>passByReference</value>
<value>exactlyOnce</value>
<value>ordered</value>
<value>py4j</value>
<value>py4j.async</value>
<value>osgi.async</value>
<value>osgi.private</value>
<value>osgi.confidential</value>
</array>
</property>
<property name="service.imported" value-type="String" value="true"/>
<property name="service.imported.configs" value-type="String">
<array>
<value>ecf.py4j.host</value>
</array>
</property>
</endpoint-description>
</endpoint-descriptions>
---End Endpoint DescriptionThis output shows the remote service was exported via Remote Service Admin and the py4j-remoteservices distribution provider with all the RSA-specified remote service meta-data in a standard format (EndpointDescription). This means that the Karaf process is now listening on Java port 25333 (default py4j port) for connections from MCP servers.
The MCP project has a testing/debugging client called the MCP Inspector. The inspector will start the {iPopo Sample Remote tools MCP Server](https://github.com/tcalmant/ipopo/tree/v3/samples/remotetoolsserver) and connect to that server using the STDIO transport (i.e. sytem in and system out) to communicate via the MCP.
See the MCP Inspector docs for instaliing and starting the inspector.
Requirements: Python 3.12+
Clone the iPOPO repo to get the sample application
git clone https://github.com/tcalmant/ipopo.git
cd ipopo
Make sure that the requirements for the sample (in samples/remotetoolsserver/requirements.txt) are
pip install -r samples/remotetoolsserver/requirements.txt
To start the inspector and have it launch (and connect to the running Java ArithmeticTools server)
set PYTHONPATH=<PARENT DIR>/ipopo;<PYTHON INSTALL DIR>/site-packages
npx @modelcontextprotocol/inspector python samples/remotetoolsserver/run_framework.py
Starting MCP inspector...
⚙️ Proxy server listening on port 6277
🔍 MCP Inspector is up and running at http://127.0.0.1:6274
With Browser then navigate to: http://127.0.0.1:6274
In the Inspector UI, click Connect, the Tools, then List Tools
Note that the Tool and ToolParam annotation content from the ArithmeticTools Java interface (for add and multiply tools) are presented in the inspector. This meta-data about the tool is dynamically added to the Python MCP Server and then communicated to the inspector via MCP.
If the Karaf/Java ArithmeticTools component is stopped in the Karaf/Java server...
karaf@root()> stop 23
Clicking List Tools in the inspector will now show that add and multiple tools have been dynamically removed from the Python MCP Server
NOTE: The inspector app currently cannot successfully call the add and multiply tools (via Call Tool button) because they are expecting integers as arguments and not strings. The inspector currently doesn't provide server specified integers input to the MCP server. Other MCP clients such as Claude, custom/test clients) do not have this problem.
There has been a new project template added to the ECF Bndtools Workspace Template that uses the ECF Python.Java Distribution Provider. This distribution provider is based upon py4j, which supports high performance remote procedure call between python and java processes.
To try it out after installing Bndtools 7.1 and the ECF tools add ons
- Create a new Bndtools Workspace using the ECF Bndtools Workspace Template
- Create a new Bnd OSGi project
- Open the projectName.hellopython.javahost.bndrun file in the project directory
-
Choose 'Resolve' and then 'Update'
-
Select Debug OSGi to start the example application (Java)
Running Python Example Program
-
Install iPOPO v 3.1.0 in your Python (3.9 or greater) local environment
-
In a command shell or IDE, navigate to the project directory and run the run_python_example.py script
python run_python_example.py
The examples will output progress to their respective consoles as the remote services are made exported, discovered, and imported by the java process or the python process.
Most of the code that produces output is available in the example project. For java: src/main/java/.../hello/*.java and python: python-src/samples/rsa
As part of recent work on iPOPO RSA, an implementation of a python import hook pep-302 was created so using remote services java bundles could resolve python packages. On the java side, the https://github.com/ECF/Py4j-RemoteServicesProvider/blob/master/bundles/org.eclipse.ecf.provider.direct/src/org/eclipse/ecf/provider/direct/ModuleResolver.java service interface exposes the methods called by a python import hook to resolve a python module.
Here is an example ModuleResolver service implementation for a python package named 'foo'. Note that the PATH_PREFIX component property points to '/python-src' inside built bundle.
To run this 'foo' package ModuleResolver in Karaf 4.4.6+
start karaf
...
# add ECF Karaf repo
karaf@root()> repo-add https://download.eclipse.org/rt/ecf/latest/karaf-features.xml
Adding feature url https://download.eclipse.org/rt/ecf/latest/karaf-features.xml
karaf@root()> feature:install ecf-rs-examples-python-importhook
(few seconds pass for download and install)
karaf@root()>
By default, a Python.Java gateway will be started and be listening for python connections on localhost:25333 (unless configured otherwise).
The 'foo' package ModuleResolver remote service will be available for resolving the 'foo' package.
The iPOPO project has a python-side Python.Java Distribution Provider, and a sample application to connect to the Java server at localhost:25333 and then resolve and run the code in the 'foo' package. This sample application is here.
To start from shell (python 3.10+):
python -m samples.run_rsa_py4j_importhook
If the Karaf server is running the python sample application will produce output like this
Attempting connect to Python.Java OSGi server listening at 25333...
** Pelix Shell prompt **
$ If the py4j localhost connect above succeeds, the code import for package foo.bar.baz
will be resolved by the OSGi server with an a active instance of a ModuleResolver
service implementation from the example in this bundle: org.eclipse.ecf.examples.importhook.module
The code for this bundle is here:
https://github.com/ECF/Py4j-RemoteServicesProvider/tree/master/examples/org.eclipse.ecf.examples.importhook.module
There are instructions for running an instance of this server in the RSA importhook tutorial at
https://ipopo.readthedocs.io/en/v3/tutorials/index.html
...importing Bar class from foo.bar.baz package
foo imported
imported bar
baz loaded
foobar loaded
...Bar class imported
...creating an instance of Bar...
Foo.init
Bar.init
...Bar instance created. The print output between the lines starting with '...' is from foo package code
The messages 'foo imported, imported bar' are produced from running the python code returned by the ModuleResolver service from /python-src.
If the Karaf server is not running/listending on localhost:25333, it will produce a connect error. Here is an example of such a connect error.
Attempting connect to Python.Java OSGi server listening at 25333...
** Pelix Shell prompt **
$ Component 'py4j-distribution-provider': error calling @ValidateComponent callback
Traceback (most recent call last):
...
py4j.protocol.Py4JNetworkError: An error occurred while trying to connect to the Java server (127.0.0.1:25333)
This example implementation HelloImpl.java shows a small OSGi Async Remote Service.
To install and start this example service in Karaf 4.4.6+ type the following into the karaf console:
karaf@root()>feature:repo-add https://download.eclipse.org/rt/ecf/latest/karaf-features.xml
karaf@root()>feature:install ecf-rs-examples-python.java-hello
After installing all bundles and components, and exporting the HelloImpl service, the Karaf console will print out the OSGi standard export registration and endpoint desription
19:20:03.679;EXPORT_REGISTRATION;exportedSR=[org.eclipse.ecf.examples.hello.IHello];cID=URIID [uri=py4j://127.0.0.1:25333/java];rsId=1
--Endpoint Description---
<endpoint-descriptions xmlns="http://www.osgi.org/xmlns/rsa/v1.0.0">
<endpoint-description<
<property name="ecf.endpoint.id" value-type="String" value="py4j://127.0.0.1:25333/java"/>
...continues with the entire endpoint description
The py4j://127.0.0.1:25333/java as the value for ecf.endpoint.id property indicates that a listener has been opened on 127.0.0.1 (localhost) using port 25333 and waiting for connections from clients.
python -m samples.run_rsa_py4j_consumer
On the Python side there should be console output
** Pelix Shell prompt **
$ Python IHello service consumer received sync response: Java says: Hi PythonSync, nice to see you
done with sayHelloAsync method
done with sayHelloPromise method
promise response: JavaPromise says: Hi PythonPromise, nice to see you
async response: JavaAsync says: Hi PythonAsync, nice to see you
On the Java/Karaf server-side should be console output
Java.sayHello called by PythonSync with message: 'Hello Java'
Java.sayHelloAsync called by PythonAsync with message: 'Hello Java'
Java.sayHelloPromise called by PythonPromise with message: 'Hello Java'```
For support please open an issue on this repo, iPOPO repo, or contact [email protected]
Python.Java Remote Services is distributed with the Apache2 license. See LICENSE in this directory for more information.