Self-hosted MCP Gateway for your private AI agents

MCPJungle is a single source-of-truth registry for all Model Context Protocol Servers running in your Organisation.

🧑‍💻 Developers use it to register & manage MCP servers and the tools they provide from a central place.

🤖 MCP Clients use it to discover and consume all these tools from a single "Gateway" MCP Server.

MCPJungle is the only MCP Server your AI agents need to connect to!

1. Developers using MCP Clients like Claude & Cursor that need to access MCP servers for tool-calling
2. Developers building production-grade AI Agents that need to access MCP servers with built-in security, privacy and Access Control.
3. Organisations wanting to view & manage all MCP client-server interactions from a central place. Hosted in their own datacenter 🔒

• Quick Start guide
• Installation
• Usage
  • Server
    • Running mcpjungle server inside Docker
    • Running mcpjungle server directly on the host machine
  • Client
    • Adding Streamable HTTP-based MCP servers
    • Adding STDIO-based MCP servers
    • Removing MCP servers
  • Connect to mcpjungle from Claude
  • Connect to mcpjungle from Cursor
  • Enabling/Disabling Tools globally
  • Tool Groups
  • Authentication
  • Enterprise features
    • Access Control
    • OpenTelemetry
• Limitations
• Contributing

This quickstart guide will show you how to:

1. Start the MCPJungle server locally using docker compose
2. Register a simple MCP server in mcpjungle
3. Connect your Claude to mcpjungle to access your MCP tools

curl -O https://raw.githubusercontent.com/mcpjungle/MCPJungle/refs/heads/main/docker-compose.yaml
docker compose up -d

Download the client binary either using brew or from the Releases.

brew install mcpjungle/mcpjungle/mcpjungle

Add the context7 remote MCP server to mcpjungle

mcpjungle register --name context7 --url https://mcp.context7.com/mcp

Add the following configuration in your Claude MCP Servers:

{
  "mcpServers": {
    "mcpjungle": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:8080/mcp",
        "--allow-http"
      ]
    }
  }
}

Try asking Claude for the following:

Use context7 to get the documentation for `/lodash/lodash`

Claude will then attempt to call the context7__get-library-docs tool via MCPJungle, which will return the documentation for the Lodash library.

Congratulations 🎉 You have successfully registered a remote MCP server in MCPJungle and called one of its tools via Claude

[!WARNING] MCPJungle is BETA software.

We're actively working to make it production-ready. You can provide your feedback by starting a discussion in this repository.

MCPJungle is shipped as a stand-alone binary.

You can either download it from the Releases Page or use Homebrew to install it:

brew install mcpjungle/mcpjungle/mcpjungle

Verify your installation by running

mcpjungle version

[!IMPORTANT] On MacOS, you will have to use homebrew because the compiled binary is not Notarized yet.

MCPJungle provides a Docker image which is useful for running the registry server (more about it later).

docker pull mcpjungle/mcpjungle

MCPJungle has a Client-Server architecture and the binary lets you run both the Server and the Client.

The MCPJungle server is responsible for managing all the MCP servers registered in it and providing a unified MCP gateway for AI Agents to discover and call tools provided by these registered servers.

The gateway itself runs over streamable http transport and is accessible at the /mcp endpoint.

For running the MCPJungle server locally, docker compose is the recommended way:

# docker-compose.yaml is optimized for individuals running mcpjungle on their local machines for personal use.
# mcpjungle will run in `development` mode by default.
curl -O https://raw.githubusercontent.com/mcpjungle/MCPJungle/refs/heads/main/docker-compose.yaml

docker compose up -d

# docker-compose.prod.yaml is optimized for orgs deploying mcpjungle on a remote server for multiple users.
# mcpjungle will run in `production` mode by default, which enables enterprise features.
curl -O https://raw.githubusercontent.com/mcpjungle/MCPJungle/refs/heads/main/docker-compose.prod.yaml

docker compose -f docker-compose.prod.yaml up -d

This will start the MCPJungle server along with a persistent Postgres database container.

You can quickly verify that the server is running:

curl http://localhost:8080/health

If you plan on registering stdio-based MCP servers that rely on npx or uvx, use mcpjungle's stdio tagged docker image instead.

MCPJUNGLE_IMAGE_TAG=latest-stdio docker compose up -d

[!NOTE] If you're using docker-compose.yaml, this is already the default image tag. You only need to specify the stdio image tag if you're using docker-compose.prod.yaml.

This image is significantly larger. But it is very convenient and recommended for running locally when you rely on stdio-based MCP servers.

For example, if you only want to register remote mcp servers like context7 and deepwiki, you can use the standard (minimal) image.

But if you also want to use stdio-based servers like filesystem, time, github, etc., you should use the stdio-tagged image instead.

[!NOTE] If your stdio servers rely on tools other than npx or uvx, you will have to create a custom docker image that includes those dependencies along with the mcpjungle binary.

Production Deployment

The default MCPJungle Docker image is very lightweight - it only contains a minimal base image and the mcpjungle binary.

It is therefore suitable and recommended for production deployments.

For the database, we recommend you deploy a separate Postgres DB cluster and supply its endpoint to mcpjungle (see Database section below).

You can see the definitions of the standard Docker image and the stdio Docker image.

You can also run the server directly on your host machine using the binary:

mcpjungle start

This starts the main registry server and MCP gateway, accessible on port 8080 by default.

The mcpjungle server relies on a database and by default, creates a SQLite DB in the current working directory.

This is okay when you're just testing things out locally.

Alternatively, you can supply a DSN for a Postgresql database to the server:

export DATABASE_URL=postgres://admin:root@localhost:5432/mcpjungle_db

#run as container
docker run mcpjungle/mcpjungle:latest

# or run directly
mcpjungle start

Once the server is up, you can use the mcpjungle CLI to interact with it.

MCPJungle currently supports MCP servers using stdio and Streamable HTTP Transports.

Let's see how to register them in mcpjungle.

Let's say you're already running a streamable http MCP server locally at http://127.0.0.1:8000/mcp which provides basic math tools like add, subtract, etc.

You can register this MCP server with MCPJungle:

mcpjungle register --name calculator --description "Provides some basic math tools" --url http://127.0.0.1:8000/mcp

If you used docker compose to run the server, and you're not on Linux, you will have to use host.docker.internal instead of your local loopback address.

mcpjungle register --name calculator --description "Provides some basic math tools" --url http://host.docker.internal:8000/mcp

The registry will now start tracking this MCP server and load its tools.

You can also provide a configuration file to register the MCP server:

cat ./calculator.json
{
  "name": "calculator",
  "transport": "streamable_http",
  "description": "Provides some basic math tools",
  "url": "http://127.0.0.1:8000/mcp"
}

mcpjungle register -c ./calculator.json

All tools provided by this server are now accessible via MCPJungle:

mcpjungle list tools

# Check tool usage
mcpjungle usage calculator__multiply

# Call a tool
mcpjungle invoke calculator__multiply --input '{"a": 100, "b": 50}'

[!NOTE] A tool in MCPJungle must be referred to by its canonical name which follows the pattern <mcp-server-name>__<tool-name>. Server name and tool name are separated by a double underscore __.

eg- If you register a MCP server github which provides a tool called git_commit, you can invoke it in MCPJungle using the name github__git_commit.

Your MCP client must also use this canonical name to call the tool via MCPJungle.

The config file format for registering a Streamable HTTP-based MCP server is:

{
  "name": "<name of your mcp server>",
  "transport": "streamable_http",
  "description": "<description>",
  "url": "<url of the mcp server>",
  "bearer_token": "<optional bearer token for authentication>"
}

Here's an example configuration file (let's call it filesystem.json) for a MCP server that uses the STDIO transport:

{
  "name": "filesystem",
  "transport": "stdio",
  "description": "filesystem mcp server",
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
}

You can register this MCP server in MCPJungle by providing the configuration file:

mcpjungle register -c ./filesystem.json

The config file format for registering a STDIO-based MCP server is:

{
  "name": "<name of your mcp server>",
  "transport": "stdio",
  "description": "<description>",
  "command": "<command to run the mcp server, eg- 'npx', 'uvx'>",
  "args": ["arguments", "to", "pass", "to", "the", "command"],
  "env": {
    "KEY": "value"
  }
}

You can also watch a quick video on How to register a STDIO-based MCP server.

[!TIP] If your STDIO server fails or throws errors for some reason, check the mcpjungle server's logs to view its stderr output.

Limitation 🚧

MCPJungle creates a new connection when a tool is called. This means a new sub-process for a STDIO mcp server is started for every tool call.

This has some performance overhead but ensures that there are no memory leaks.

But it also means that currently MCPJungle doesn't support stateful connections with your MCP server.

We want to hear your feedback to improve this mechanism, feel free to create an issue, start a discussion or just reach out on Discord.

You can remove a MCP server from mcpjungle.

mcpjungle deregister calculator
mcpjungle deregister filesystem

Once removed, this mcp server and its tools are no longer available to you or your MCP clients.

Assuming that MCPJungle is running on http://localhost:8080, use the following configurations to connect to it:

{
  "mcpServers": {
    "mcpjungle": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:8080/mcp",
        "--allow-http"
      ]
    }
  }
}

{
  "mcpServers": {
    "mcpjungle": {
      "url": "http://localhost:8080/mcp"
    }
  }
}

You can watch a quick video on How to connect Cursor to MCPJungle.

You can enable or disable a specific tool or all the tools provided by an MCP Server.

If a tool is disabled, it is not available via the MCPJungle Proxy, so no MCP clients can view or call it.

# disable the `get-library-docs` tool provided by the `context7` MCP server
mcpjungle disable context7__get-library-docs

# re-enable the tool
mcpjungle enable context7__get-library-docs

# disable all tools provided by the `context7` MCP server
mcpjungle disable context7

# re-enable all tools of `context7`
mcpjungle enable context7

A disabled tool is still accessible via mcpjungle's HTTP API, so humans can still manage it from the CLI (or any other HTTP client).

[!NOTE] When a new server is registered in MCPJungle, all its tools are enabled by default.

As you add more MCP servers to MCPJungle, the number of tools available through the Gateway can grow significantly.

If your MCP client is exposed to hundreds of tools through the gateway MCP, its performance may degrade.

MCPJungle allows you to expose only a subset of all available tools to your MCP clients using Tool Groups.

You can create a new group and only include specific tools that you wish to expose.

Once a group is created, mcpjungle returns a unique endpoint for it.

You can then configure your MCP client to use this group-specific endpoint instead of the main gateway endpoint.

You can create a new tool group by providing a JSON configuration file to the create group command.

You must specify a unique name for the group and a list of included_tools that you want to expose via its MCP proxy.

Here is an example of a tool group configuration file (claude-tools-group.json):

{
  "name": "claude-tools",
  "description": "This group only contains tools for Claude Desktop to use",
  "included_tools": [
    "filesystem__read_file",
    "deepwiki__read_wiki_contents",
    "time__get_current_time"
  ]
}

Instead of exposing 20 tools across all MCP servers, this group only exposes 3 handpicked ones.

You can create this group in mcpjungle:

$ mcpjungle create group -c ./claude-tools-group.json

Tool Group claude-tools created successfully
It is now accessible at the following streamable http endpoint:

    http://127.0.0.1:8080/v0/groups/claude-tools/mcp

You can then configure Claude (or any other MCP client) to use this group-specific endpoint to access the MCP server.

The client will then ONLY see and be able to use these 3 tools and will not be aware of any other tools registered in MCPJungle.

[!TIP] You can run mcpjungle list tools to view all available tools and pick the ones you want to include in your group.

You can also watch a Video on using Tool Groups.

You can currently perform operations like listing all groups, viewing details of a specific group and deleting a group.

# list all tool groups
mcpjungle list groups

# view details of a specific group
mcpjungle get group claude-tools

# delete a group
mcpjungle delete group claude-tools

[!NOTE] If a tool is included in a group but is later disabled globally or deleted, then it will not be available via the group's MCP endpoint.

But if the tool is re-enabled or added again later, it will automatically become available in the group again.

Limitations 🚧

1. Currently, you cannot update an existing tool group. You must delete the group and create a new one with the modified configuration file.
2. In production mode, currently only an admin can create a Tool Group. We're working on allowing standard Users to create their own groups as well.

MCPJungle currently supports authentication if your Streamable HTTP MCP Server accepts static tokens for auth.

This is useful when using SaaS-provided MCP Servers like HuggingFace, Stripe, etc. which require your API token for authentication.

You can supply your token while registering the MCP server:

# If you specify the `--bearer-token` flag, MCPJungle will add the `Authorization: Bearer <token>` header to all requests made to this MCP server.
mcpjungle register --name huggingface --description "HuggingFace MCP Server" --url https://huggingface.co/mcp --bearer-token <your-hf-api-token>

Or from your configuration file

{
  "name": "huggingface",
  "transport": "streamable_http",
  "url": "https://huggingface.co/mcp",
  "description": "hugging face mcp server",
  "bearer_token": "<your-hf-api-token>"
}

Support for Oauth flow is coming soon!

If you're running MCPJungle in your organisation, we recommend running the Server in the production mode:

# enable enterprise features by running in production mode
mcpjungle start --prod

# you can also specify the server mode as environment variable (valid values are `development` and `production`)
export SERVER_MODE=production
mcpjungle start

# Or use the production docker compose file as described above
docker compose -f docker-compose.prod.yaml up -d

By default, mcpjungle server runs in development mode which is ideal for individuals running it locally.

In Production mode, the server enforces stricter security policies and will provide additional features like Authentication, ACLs, observability and more.

After starting the server in production mode, you must initialize it by running the following command on your client machine:

mcpjungle init-server

This will create an admin user in the server and store its API access token in your home directory (~/.mcpjungle.conf).

You can then use the mcpjungle cli to make authenticated requests to the server.

In development mode, all MCP clients have full access to all the MCP servers registered in MCPJungle Proxy.

production mode lets you control which MCP clients can access which MCP servers.

Suppose you have registered 2 MCP servers calculator and github in MCPJungle in production mode.

By default, no MCP client can access these servers. You must create an MCP Client in mcpjungle and explicitly allow it to access the MCP servers.

# Create a new MCP client for your Cursor IDE to use. It can access the calculator and github MCP servers
mcpjungle create mcp-client cursor-local --allow "calculator, github"

MCP client 'cursor-local' created successfully!
Servers accessible: calculator,github

Access token: 1YHf2LwE1LXtp5lW_vM-gmdYHlPHdqwnILitBhXE4Aw
Send this token in the `Authorization: Bearer {token}` HTTP header.

Mcpjungle creates an access token for your client. Configure your client or agent to send this token in the Authorization header when making requests to the mcpjungle proxy.

For example, you can add the following configuration in Cursor to connect to MCPJungle:

{
  "mcpServers": {
    "mcpjungle": {
      "url": "http://localhost:8080/mcp",
      "headers": {
        "Authorization": "Bearer 1YHf2LwE1LXtp5lW_vM-gmdYHlPHdqwnILitBhXE4Aw"
      }
    }
  }
}

A client that has access to a particular server this way can view and call all the tools provided by that server.

[!NOTE] If you don't specify the --allow flag, the MCP client will not be able to access any MCP servers.

MCPJungle supports Prometheus-compatible OpenTelemetry Metrics for observability.

• In production mode, OpenTelemetry is enabled by default.
• In development mode, telemetry is disabled by default. You can enable it by setting the OTEL_ENABLED environment variable to true before starting the server:

# enable OpenTelemetry metrics
export OTEL_ENABLED=true

# optionally, set additional attributes to be added to all metrics
export OTEL_RESOURCE_ATTRIBUTES=deployment.environment.name=production

# start the server
mcpjungle start

Once the mcpjungle server is started, metrics are available at the /metrics endpoint.

We're not perfect yet, but we're working hard to get there!

When you call a tool in a Streamable HTTP server, mcpjungle creates a new connection to the server to serve the request.

When you call a tool in a STDIO server, mcpjungle creates a new connection and starts a new sub-process to run this server.

After servicing your request, it terminates this sub-process.

So a new stdio server process is started for every tool call.

This has some performance overhead but ensures that there are no memory leaks.

It also means that if you rely on stateful connections with your MCP server, mcpjungle can currently not provide that.

We plan on improving this mechanism in future releases and are open to ideas from the community!

This is a work in progress.

We're collecting more feedback on how people use OAuth with MCP servers, so feel free to start a Discussion or open an issue to share your use case.

We welcome contributions from the community!

• For contribution guidelines and standards, see CONTRIBUTION.md
• For development setup and technical details, see DEVELOPMENT.md

Join our Discord community to connect with other contributors and maintainers.