minibridge

# Minibridge Minibridge serves as a backend-to-frontend bridge, streamlining and securing communication between Agents and MCP servers. It safely exposes MCP servers to the internet and can optionally integrate with generic policing services — known as Policers — for agent authentication, content analysis, and transformation. Policers can be implemented remotely via HTTP or locally using OPA Rego policies. Additionally, Minibridge can help ensure the integrity of MCP servers through SBOM (Software Bill of Materials) generation and real-time validation. ## Table of Content <!-- vim-markdown-toc GFM --> * [All In One](#all-in-one) * [Backend](#backend) * [Frontend](#frontend) * [Stdio](#stdio) * [HTTP+SSE](#httpsse) * [MCP Server Integrity Check](#mcp-server-integrity-check) * [Generate SBOM](#generate-sbom) * [Check SBOM Offline](#check-sbom-offline) * [Check SBOM Online](#check-sbom-online) * [Dump everything](#dump-everything) * [Policer](#policer) * [Policer API](#policer-api) * [Police Request](#police-request) * [Police Response](#police-response) * [Available Policers](#available-policers) * [HTTP Policer](#http-policer) * [Rego Policer](#rego-policer) * [Agent Authentication](#agent-authentication) * [Global](#global) * [Forward](#forward) * [Todos](#todos) <!-- vim-markdown-toc --> ## All In One Minibridge can act as a single gateway positioned in front of a standard stdio-based MCP server. To start everything as a single process, run: minibridge aio --listen :8000 -- npx @modelcontextprotocol/server-filesystem /tmp This command launches both the frontend and backend within a single process, which can be useful in certain scenarios. You can connect directly using an HTTP client: $ curl http://127.0.0.1:8000/sse event: endpoint data: /message?sessionId=UID $ curl http://127.0.0.1:8000/message?sessionId=UID \ -X POST \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' The flow will look like the following: ```mermaid flowchart LR agent -- http+sse --> minibridge minibridge -- stdio --> mcpserver ``` In order to secure the connections, you need to enable HTTPS for incoming connections: minibridge aio --listen :8443 \ --tls-server-cert ./server-cert.pem \ --tls-server-key ./server-key.pem \ --tls-server-client-ca ./clients-ca.pem \ -- npx @modelcontextprotocol/server-filesystem /tmp This enables HTTPS and with `--tls-server-client-ca`, it requires the clients to send a certificate signed by that client CA. You can now connect directly using an HTTP client: $ curl https://127.0.0.1:8443/sse \ --cacert ./server-cert.pem --cert ./client-cert.pem --key ./client-key.pem event: endpoint data: /message?sessionId=UID $ curl https://127.0.0.1:8443/message?sessionId=UID \ --cacert ./server-cert.pem --cert ./client-cert.pem --key ./client-key.pem \ -X POST \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' ## Backend Starting the backend launches an MCP server and exposes its API over a WebSocket-based interface. TLS can be configured, with or without client certificates, depending on your security requirements. For example, to start a filesystem-based MCP server: minibridge backend -- npx @modelcontextprotocol/server-filesystem /tmp You can now connect directly using a websocket client: wscat --connect ws://127.0.0.1:8000/ws > NOTE: use the `wss` scheme if you have started Minibridge backend with TLS. > NOTE: Today, Minibridge backend only supports MCP server over stdio. The flow will look like the following: ```mermaid flowchart LR agent -- websocket --> minibridge minibridge -- stdio --> mcpserver ``` In order to secure the connections, you need to enable HTTPS for incoming connections: minibridge backend --listen :8443 \ --tls-server-cert ./backend-server-cert.pem \ --tls-server-key ./backend-server-key.pem \ --tls-server-client-ca ./clients-ca.pem \ -- npx @modelcontextprotocol/server-filesystem /tmp This enables HTTPS and with `--tls-server-client-ca`, it requires the clients to send a certificate signed by that client CA. You can now connect using: wscat --connect wss://127.0.0.1:8443/ws \ --ca ./server-cert.pem \ --cert ./client-cert.pem ## Frontend While WebSockets address many of the limitations of plain POST+SSE, they are not yet part of the official MCP protocol. To maintain backward compatibility with existing agents, the frontend can expose a local interface using POST+SSE, HTTP+STREAM (coming soon), or plain STDIO. It will then transparently forward the data to the Mminibridge backend over WebSockets and HTTPS. ### Stdio To start an stdio frontend: minibridge frontend --backend wss://127.0.0.1:8000/ws You can then send requests via stdin and read responses from stdout. The frontend maintains a single WebSocket connection to the backend and will automatically reconnect in case of failures. The flow will look like the following: ```mermaid flowchart LR agent -- stdio --> mb1[minibridge] mb1 -- websocket --> mb2[minibridge] mb2 -- stdio --> mcpserver ``` ### HTTP+SSE To start an SSE frontend: minibridge frontend --listen :8081 --backend wss://127.0.0.1:8000/ws In this mode, a new WebSocket connection is established with the backend for each incoming connection to the /sse endpoint. This preserves session state. However, the WebSocket will not attempt to reconnect in this mode, and any active streams will be terminated in the event of a network failure. You can connect directly using an HTTP client: $ curl http://127.0.0.1:8001/sse event: endpoint data: /message?sessionId=UID $ curl http://127.0.0.1:8001/message?sessionId=UID \ -X POST \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' The flow will look like the following: ```mermaid flowchart LR agent -- http+sse --> mb1[minibridge] mb1 -- websocket --> mb2[minibridge] mb2 -- stdio --> mcpserver ``` In order to secure the connections, you need to enable HTTPS for incoming connections: minibridge frontend --listen :8444 \ --backend wss://127.0.0.1:8000/ws --tls-server-cert ./server-cert.pem \ --tls-server-key ./server-key.pem \ --tls-server-client-ca client-ca.pem ]\ --tls-client-cert ./client-cert.pem \ --tls-client-key ./client-key.pem \ --tls-client-backend-ca ./backend-server-cert.pem This enables HTTPS and with `--tls-server-client-ca`, it requires the clients to send a certificate signed by that client CA. It also make the front end to authenticate to the backend using the provided client certificate (MTLS). You can now connect directly using an HTTP client: $ curl https://127.0.0.1:8444/sse \ --cacert ./server-cert.pem --cert ./client-cert.pem --key ./client-key.pem event: endpoint data: /message?sessionId=UID $ curl https://127.0.0.1:8444/message?sessionId=UID \ --cacert ./server-cert.pem --cert ./client-cert.pem --key ./client-key.pem \ -X POST \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' ## MCP Server Integrity Check Minibridge can generate an SBOM file for a MCP server that contains hashes of various parts interpreted by an LLM. This SBOM file can be used to check the MCP Server did not change, either offline or online by passing it to the backend that will inspect each relevant MCP call dynamically. ### Generate SBOM To generate a SBOM file: minibridge scan sbom -- npx @modelcontextprotocol/server-everything > server.sbom This will create `server.sbom` that contains all the hashes of the relevant parts. > NOTE: For now, only tools and prompts are part of the SBOM. ### Check SBOM Offline To verify that an MCP server still matches a previously generated SBOM file: minibridge scan check server.sbom -- npx @modelcontextprotocol/server-everything ### Check SBOM Online The SBOM file can be passed to the Minibridge backend to dynamically ensure no LLM-interpreted parts have changed since the SBOM file generation. For instance, it can make sure a tool that reads a file does not suddenly change to a tool that sends all your credentials to some shady place. To start Minibridge backend with live SBOM validation: minibridge backend -l :8000 --sbom server.sbom -- npx @modelcontextprotocol/server-everything ### Dump everything Dump all the tools, prompts, resources and resource templates of a particular MCP Server. For example: minibridge scan dump -- npx @modelcontextprotocol/server-everything ## Policer While Minibridge already offers advanced features such as strong client authentication and native WebSocket support, it can be further enhanced through integration with a Policer. A Policer is responsible for: * Authentication * Authorization * Input analysis and logging * Full request tracing * And more advanced policy-based controls There are various available policers in Minibridge: - HTTP policer: sends a request to a remote HTTP service to delegate decision - Rego policer: runs a [rego](https://www.openpolicyagent.org/docs/latest) policy file on the request > NOTE: More policers will be added! The Policer, if set, will be called and passed various information so it can make a decision on what to do with the request, based on the user who initiated the request and the content of the request. You can then start Minibridge, using either the aio or backend subcommand, with the following arguments to start an HTTP policer: minibridge aio --policer-type http \ --policer-http-url https://policer.acme.com/police --policer-http-token $PTOKEN Or using with the following command to start Minibridge with a Rego policer: minibridge aio --policer-type rego \ --policer-rego-policy ./example/policer-rego/policy.rego Once integrated, any command from the user or response from the MCP Server received by the backend is first passed to the Policer for authentication and/or analysis. ### Policer API #### Police Request The Policer will receives the following information: ```json { "type": "request|response" "agent": { "token": "<agent-token>", "userAgent": "curl/7.54.1", "remoteAddr": "10.0.1.265:44536", }, "mcp": { "jsonrpc": "2.0", "id": 1, "method": "tools/list", } } ``` The Policer can use this information to decide if the request should be denied. #### Police Response The policer must return a response with the following information: ```json { "deny": ["reason 1", "reason 2"], "mcp": null, } ``` If `deny` is set an not empty, then the request or response is considered as blocked and Minibridge will not forward it to the MCP server or to the Agent. Instead, it will return a descriptive MCP error containing the deny reasons, to the caller. If `deny` is `null` or empty, the call is allowed and will be forwarded. In addition, if the call is allowed, and `mcp` is non null, Minibridge will swap the original MCP call with the one provider in the response, allowing Policers to mutate the call. For instance this can be used to hide some tools based on the agent identity. ### Available Policers #### HTTP Policer The HTTP Policer will receive the Police Request as `POST` on the url provided by `--provider-http-url`. To allow the request or response, the HTTP Policer must respond with: - An HTTP status `204 No Content`, or - An HTTP status `200 OK` with `deny` property set to `null` or `[]` To disallow the request or response: - An HTTP status `200 OK` in case of deny with a valid response. - Any other HTTP code will also deny the request, but is considered as an error. For example, a policy result that denies the request: ```http HTTP/1.1 200 OK { "deny": ["You are not allowed to list the tools"] } ``` And a policy result that permits the request: ```http HTTP/1.1 204 No Content ``` Or ```http HTTP/1.1 200 OK { "deny": [] } ``` The HTTP Policer can also decide to mutate the MCP call. To do so, it must allow the request, and pass back a modified MCP call: ```http HTTP/1.1 200 OK { "mcp": { "id": 2, "jsonrpc": "2.0", "result": { "tools": [{ "description": "POLICER HAS MODIFIED THIS DESCRIPTION", "name": "echo" }] } } ``` #### Rego Policer The Rego Policer allows to run the Police Request through a rego policy to decide if the request should be allowed or not. The Police Request is passed an input, and the rego policy must either return and `allow := true` or an empty `deny` string array. > NOTE: the rego package must be named `main`. For instance, to allow the request: ```regp package main import rego.v1 allow := true ``` To deny the request: ```rego package main import rego.v1 deny contains msg if { input.agent.token == "" msg := "you must send a token" } deny contains msg if { input.mcp.method == "tools/call" input.mcp.params.name == "printEnv" msg := "you cannot use printEnv" } ``` ### Agent Authentication The frontend can forward the agent credentials using two modes. #### Global This mode uses a single token set during the Minibridge frontend startup. It is useful when there is only one user of the bridge — such as when it is running on a user’s laptop and connecting to a remote Minibridge backend. To start the frontend in that mode: minibridge frontend -l :8000 -A wss://backend.minibridge.acme.com/ws --agent-token "$TOKEN" #### Forward This mode simply forwards the HTTP `Authorization` header to the Minibridge backend. It is useful when the frontend runs on a public network and you want the policer to authenticate multiple callers. To start the frontend in that mode: minibridge frontend -l :8000 -A wss://backend.minibridge.acme.com/ws --agent-token-passthrough ## Todos Minibridge is still missing the following features: - [ ] Unit tests - [x] Transport user information over the websocket channel - [x] Support for user extraction to pass to the policer - [ ] Optimize communications between front/back in aio mode - [ ] Plug in prometheus metrics