UplinkCrate
Connect your AI clients to your tools
Open-source, self-hosted MCP gateway.
Give compatible AI clients authenticated access to multiple Model Context Protocol (MCP) servers through one gateway. Powered by MCPHub and Caddy, with Streamable HTTP and SSE transports.
Open source · MIT licensed
The Self-Hosted AI Builder's Guide Practical guidance for deployment, security, backups, and everyday operation.
AI client → gateway → tools
/mcp · Streamable HTTP
/mcp/fetch · additional servers you configure
A focused service for your applications
Connect applications to the tool servers you configure, with gateway authentication and a central place to manage server connections. Keep permissions and approval rules in your client as well.
Connect through the aggregated /mcp endpoint or a per-server route such as /mcp/fetch. Your client must support the chosen transport.
Expose compatible MCP servers to HTTP clients using Streamable HTTP or SSE. This is an MCP gateway, not a separate LLM inference service.
Caddy protects API access with a Bearer token. Fresh starts generate and store a key in the data volume; dashboard administrator credentials are managed separately.
Packaged integrations include fetch, filesystem, GitHub, Brave Search, Git, PostgreSQL, memory, and sequential thinking. Enable and configure the servers you need.
Start with an env file and manage persistent server configuration through MCPHub. Dashboard access also needs the gateway’s authentication; see the setup guide.
Run on amd64 or arm64 Linux hosts. Persistent data retains gateway configuration and credentials. No GPU is required for the gateway itself.
Images are automatically built and published through public GitHub Actions workflows. Powered by MCPHub, Caddy.
Start with the fetch server
This example binds the gateway to the Docker host. Fetch is enabled by default; additional integrations may need credentials, mounts, or configuration.
For access from other computers, follow the maintained HTTPS, authentication, and network-exposure instructions.
docker run \
--name mcp \
--restart=always \
-v mcp-data:/var/lib/mcp \
-p 127.0.0.1:3000:3000 \
-d hwdsl2/mcp-gatewayCheck initial startup with docker logs mcp before retrieving the key.
uplink_api_key="$(docker exec mcp mcp_manage --getkey)"Fresh starts generate a Bearer key and save it in persistent data. Keep this key private.
curl http://127.0.0.1:3000/mcp \
-H "Authorization: Bearer $uplink_api_key" \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"quickstart","version":"1.0"}}}'This is an authenticated protocol check. For actual tool use, configure your MCP client with this endpoint and Bearer header and follow its session-handling instructions.
Part of a larger toolkit
UplinkCrate is also available as part of Self-Hosted AI Stack. Deploy local models, chat, document processing, voice, and AI tools together, or choose a smaller configuration for your needs.
Explore Self-Hosted AI StackConnect the services you need
GatewayCrate can route model requests and connect compatible model workflows to MCP tools. Configure both services and keep tool permissions explicit.
Explore GatewayCrateThe UplinkCrate project is open source under the MIT License. You provide hardware or cloud hosting and pay associated costs. Models, dependencies, and external providers retain their own licenses and terms.
No. It connects MCP clients to tool servers. A client or agent runtime supplies the model, decides which tools to call, and enforces its own approvals and stopping rules.
The fetch server is enabled by default. Other packaged integrations require the relevant configuration, credentials, or mounted paths. Availability does not grant permission to access arbitrary files or accounts.
No. Fetch, search, GitHub, databases, and other integrations can contact external systems. Check destinations, credentials, and permissions for each server and client.
Mount only the paths a tool needs, use read-only access when appropriate, and restrict the server’s allowed directories. Avoid exposing your whole host filesystem.
The dashboard sits behind gateway Bearer authentication and has its own administrator credentials. Follow the documented authentication setup rather than assuming a plain browser visit is sufficient.
The image includes optional aggregate usage counts.
See the usage-count documentation;
set MCP_DISABLE_USAGE_COUNTS=1 to disable them.
Connect the tools you need. Keep their permissions deliberate.
View on GitHub