Solace Broker MCP Server: Get an AI Assist Managing Your Brokers

Ask an AI assistant why a queue is backing up and what you don’t want is a generic answer about consumers and flow control. Ask it with the Solace Broker MCP Server connected and it can look at the queue, check the clients bound to it, see the discards, and tell you what is actually happening on your broker.

The Solace Broker MCP Server (v1.0.0) connects any MCP-compatible client, to your Solace event brokers, so operators and developers can monitor, diagnose and manage them in plain language instead of memorizing SEMP syntax or hopping between the CLI, Broker Manager and REST tooling.

Why Didn’t we Just Expose the API?

The fastest way to build an MCP server is to generate one tool per API endpoint. It takes very little code and gives you complete coverage on day one. We looked at that option, ran it by some customers, and decided against it for reasons of praticallity and usefulness.

Every tool you expose has to be described to the model, and those descriptions land in the context window before the assistant has done any work. Hundreds of tools means hundreds of descriptions, higher token costs, and more chances for the model to pick the wrong one or invent one that does not exist. Early feedback on our first prototype said just that: it was slow, and it hallucinated tools and APIs.

So the Broker MCP Server is a curated set of 47 tools, chosen around the things operators do every day. It is deliberately not the full SEMP API. A few things follow from that:

• A small context footprint. The assistant sees a focused toolbox, and the server can narrow tools/list further so a caller only sees the tools their group is allowed to use.
• Tools shaped like operator tasks. You get queue depth, spool usage and rates, slow-subscriber detection, bridge failure reasons and RDP binding state, not a pile of raw endpoints to stitch together.
• A light touch on the broker. The server acts as a well-behaved SEMP client, with per-broker rate limiting, retry and backoff, and concurrency caps. Your assistant should not become the noisiest thing talking to your production event brokers and event mesh.
• A smaller attack surface. Fewer tools means fewer things to secure, audit and explain to your security team.

If you need an endpoint we have not covered, the SEMP API is still there. The MCP server is built for the high-frequency jobs, and we will expand coverage based on what customers ask for rather than starting at 200+ tools and trimming back. Customers can do this easily through the ideas portal here – Introducing the Solace Ideas Portal — your feature requests with full visibility into the progress – General Discussions – Solace Community

What do Broker MCP Server’s 47 Tools Cover?

There are 29 read-only tools and 18 write tools.

• The Read Only Tools: The read tools cover broker health and redundancy, replication, Message VPNs, queues, clients, client usernames and profiles, REST Delivery Points, bridges, Kafka receivers and senders, and discard diagnostics. There is also a schema introspection tool that lets the assistant look up the live SEMP schema for an operation when it needs to. Client username and profile inspection is read-only and never returns secrets.
• Write Capable Tools: The write tools disconnect clients, purge queues, reset statistics, and create, update and delete Message VPNs, queues, queue subscriptions, topic endpoints and REST Delivery Points. All of them are off by default. An operator has to opt in explicitly with enable_write_tools, so a team can start with read-only diagnostics and turn on changes when they are ready.

A single server instance can handle many brokers, each addressed by an alias you configure, so a fleet does not need a fleet of MCP servers.

Built to Satisfy Security and Audit Teams

Giving an AI assistant access to infrastructure raises fair questions, so most of the 1.0 work went into the boring parts (with respect to our peers in security and compliance!).

• Real identity, not a shared bot account. In production the server supports OAuth 2.1 and OIDC, validated against Keycloak and Microsoft Entra ID. It can pass the calling user’s identity through to the broker using token exchange or Entra On-Behalf-Of, so the broker’s audit log records who actually made the request.
• Authorization at the tool level. You can gate individual tools by a group or role claim in the caller’s token. An on-call engineer might get purge and disconnect, while a developer gets read-only.
• Secrets stay away from the model. Credentials live in configuration only, are redacted from logs, and are never exposed to the LLM.
• Every destructive action leaves a trail. There is an audit log for every destructive operation, and a correlation ID on every request that travels from the agent through the MCP server to the broker.

Easy to Run, Easy to Watch

The server ships as a single Go binary, a Docker image, or via go install, with Kubernetes manifests included. It holds no persistent state, so upgrading means starting a new instance and letting clients reconnect. It works with Solace event brokers, software and appliance, on both LTS and rolling releases currently under Solace support, with no coupling to a specific broker version.

For operations teams, it exposes Prometheus metrics and OpenTelemetry tracing, Kubernetes liveness and readiness probes, an importable Grafana dashboard and an operator runbook. It is licensed under Apache 2.0.

Pugged into Solace Agent Mesh, you can wire the server into an Agent Mesh project as an MCP-tooled agent.

How Does this Compares with other MCP Servers?

MCP servers for messaging and streaming platforms are appearing quickly, and they make very different trade-offs. Some are open source samples with a couple of tools, such as one that simply passes administrative commands through to the broker, which is flexible but leaves safety and structure to the user. Some are read-only agents that diagnose problems but cannot act on them, or are limited in how many brokers they can cover at once. Others offer 50 or more tools spread across several services, are community supported on a best-effort basis, and rely on the underlying platform permissions for access control.

We aimed for a different balance: enough tools to be useful for real operations, few enough to keep the model focused, and identity, authorization and audit built in from the start.

Get Started

Point your MCP client at the server and ask something real. Try:

• “List all my brokers.”
• “What version is the lab broker running?”
• “Which queues on this VPN are backing up, and which clients are bound to them?”
• “Show me discard diagnostics for the orders VPN.”

Setup is documented for Claude Code and Claude Desktop, and any MCP client will work.

• Repository: https://github.com/SolaceProducts/solace-broker-mcp
• v1.0.0 release: https://github.com/SolaceProducts/solace-broker-mcp/releases/tag/v1.0.0
• Tools reference: https://github.com/SolaceProducts/solace-broker-mcp/blob/main/docs/tools-reference.md
• User guide: https://github.com/SolaceProducts/solace-broker-mcp/blob/main/docs/user-guide.md

If you are upgrading from a pre-GA release, read the changelog first. One change to know about: broker traffic now follows the standard proxy environment variables where it previously went direct, so exempt your brokers (and your OTLP collector, if telemetry is on) before you upgrade.