muster MCP server
Commit Nov 14, 2026Aggregates multiple MCP servers behind one authenticated endpoint so AI clients discover and call their tools through meta-tools.
Also known as giantswarm/muster
- Stars
- 40
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 14, 2026
- License
- Apache-2.0
What is muster MCP?
muster aggregates multiple MCP servers behind one authenticated endpoint so AI clients discover and call their tools through meta-tools.
Summary generated from the project's README
About muster MCP server
Aggregating various MCP servers into a single authenticated endpoint is what muster enables for organizations. It allows teams to connect diverse systems like Kubernetes, Prometheus, or internal service desks and serve their capabilities through one gateway. Instead of loading hundreds of tool definitions into a context window, AI clients use meta-tools to discover and call specific functions.
Developers can run muster as a single binary on a laptop to bridge an IDE to local or remote servers, or deploy it as a Kubernetes service for organization-wide access. It manages authentication via OAuth 2.1 and Dex, ensuring that downstream calls are made using the identity of the person working. The system uses toolsets to limit the visible surface area for specific agents and supports deterministic workflows for multi-step procedures. It provides observability through OpenTelemetry traces and Prometheus metrics to monitor every session and tool call.
What you can do with it
- 1Consolidate diverse tool sourcesAn organization can connect its CRM, ERP, and maintenance systems into one unified interface for AI agents.
- 2Optimize context window usageA developer can use meta-tools to find and describe specific tools without overwhelming the AI with a massive tool catalogue.
- 3Enforce identity-based accessEnsure that every tool execution performed by an agent is authenticated under the specific user's identity.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablegiantswarm/muster is public and reachablegithub.com/giantswarm/muster ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)Apache-2.0spdx.org/licenses/Apache-2.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/giantswarm/muster ↗Pass
- OpenSSF ScorecardOpenSSF Scorecard 5.9 / 10; weakest: Code-Review 0, Token-Permissions 0, CII-Best-Practices 0scorecard.dev/viewer/?uri=github.com/giantswarm/muster ↗Partial 9/15
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubmuster is an aggregating Model Context Protocol server. It connects to the MCP servers an organisation runs (a factory's MES and maintenance system, a service desk and its directory, a CRM and an ERP, Kubernetes and Prometheus, in-house tools) and serves all of their tools through a single MCP endpoint. AI agents connect once, discover tools through a small set of meta-tools instead of loading hundreds of tool definitions, and call them under the identity of the person they work for.
It runs as a single binary on a laptop, bridging an IDE to local and remote MCP servers, and as a Kubernetes service that gives a whole organisation one authenticated, observable tool gateway.
What it does
- Aggregation. Registered MCP servers (
stdio,streamable-http,sse) are connected, health-checked and reconnected; their tools appear asx_<server>_<tool>. Servers that offer the same tools for different targets can form a family and appear as one tool surface with an instance argument. - Discovery that fits a context window. Clients see thirteen meta-tools:
list_tools,filter_tools,describe_tool,call_tooland their resource and prompt counterparts. Listings are paged and summarised;filter_toolsranks by relevance, so finding the right tool costs a few hundred tokens rather than the whole catalogue. - Toolsets. A request declares the tools it may see and call with the
X-muster-Toolsetheader: built-in presets such asread-only, operator-defined presets, and inline selectors by server, pattern, workflow or label. An agent runs with exactly the surface it was given. - Authentication and single sign-on. muster protects its endpoint as an OAuth 2.1 resource server in front of Dex, keeps one session per person and forwards the person's identity token to the MCP servers that accept it, so every downstream call is made as that person. Remote servers with their own OAuth get a browser login once per person; AWS-hosted servers can be signed with SigV4.
- Kubernetes-native.
MCPServer,WorkflowandWorkflowExecutionare custom resources reconciled by muster, with status, conditions and events. The Helm chart ships the CRDs, RBAC, network policies, a Prometheus rule set and a Grafana dashboard. - Workflows. Multi-step procedures become deterministic tools (
workflow_<name>) with templated arguments, conditions,forEachloops (nested, with bounded parallelism) andparallelsteps, paged listings,jqresult projection and a durable execution record. - Observability. OpenTelemetry traces, Prometheus metrics and structured logs for every session, server and tool call.
How it works
flowchart LR
subgraph clients [MCP clients]
IDE["IDE or coding agent<br/>(Cursor, VS Code, Claude Code)"]
Agent["Platform agents<br/>(kagent, Backstage, custom)"]
end
subgraph muster [muster]
Meta["Meta-tools<br/>list_tools · filter_tools · describe_tool · call_tool"]
Auth["OAuth 2.1 resource server<br/>sessions · toolsets · SSO"]
Agg["Aggregator<br/>registry · health · reconnect"]
end
subgraph servers [MCP servers]
K8s["mcp-kubernetes"]
Prom["mcp-prometheus"]
GH["GitHub MCP"]
Own["your own"]
end
Dex["Dex (OIDC)"]
IDE -- "MCP over HTTP" --> Auth
Agent -- "MCP over HTTP" --> Auth
Auth --> Meta --> Agg
Auth -. "login, token validation" .-> Dex
Agg -- "identity token forwarded" --> K8s
Agg --> Prom
Agg --> GH
Agg --> Own
An MCP client connects to /mcp and receives the meta-tools. filter_tools and describe_tool
find and explain a tool; call_tool runs it. muster resolves the name to the server behind it,
attaches the session's credentials and forwards the call.
Quick start
Install the latest release. Binaries are signed in CI and verified by muster self-update.
brew trust giantswarm/muster
brew install giantswarm/muster/muster
Without Homebrew, download the binary for the platform:
os="$(uname -s | tr '[:upper:]' '[:lower:]')"
arch="$(uname -m | sed 's/x86_64/amd64/; s/aarch64/arm64/')"
curl -fsSL -o muster "https://github.com/giantswarm/muster/releases/latest/download/muster-${os}-${arch}"
chmod +x muster && sudo mv muster /usr/local/bin/
With a Go toolchain, go install github.com/giantswarm/muster/v5@latest works as well.
Start the aggregator, register an MCP server and call one of its tools:
muster serve # serves http://localhost:8090/mcp
# in a second terminal
muster create mcpserver files --type=stdio --command=npx \
--args="-y,@modelcontextprotocol/server-filesystem,$HOME" --autoStart=true
muster list mcpserver # files Running stdio
muster call x_files_list_allowed_directories
Connect an IDE. muster standalone runs the aggregator and a stdio bridge in one process; with a
running muster serve, or a remote muster, use muster agent --mcp-server instead.
{
"mcpServers": {
"muster": {
"command": "muster",
"args": ["standalone"]
}
}
}
A remote muster protected by OAuth asks for a sign-in first. muster auth login --endpoint <url>
opens a browser (the commands in $BROWSER before the platform default) and stores the session
once the browser returns to http://localhost:3000/callback. Where the CLI cannot reach a browser
(a remote shell, an agent's session), muster auth login --no-browser prints the sign-in URL to
open by hand in a browser on the same machine and waits for the callback. Authenticate the
CLI covers both.
The quick start continues from here: exploring the catalogue in the REPL, toolsets and the first workflow.
Deploy on Kubernetes
The chart is published in the Giant Swarm catalog and runs muster in Kubernetes mode, where MCP servers and workflows are custom resources.
helm repo add giantswarm https://giantswarm.github.io/giantswarm-catalog/
helm install muster giantswarm/muster --namespace muster --create-namespace
apiVersion: muster.giantswarm.io/v1alpha1
kind: MCPServer
metadata:
name: kubernetes
namespace: muster
spec:
type: streamable-http
url: https://mcp-kubernetes.example.com/mcp
auth:
type: oauth
forwardToken: true
The installation guide covers OAuth protection with Dex, Valkey-backed sessions for more than one replica, network policies and metrics.
Documentation
The documentation is published at giantswarm.github.io/muster
and lives in docs/; the site's landing page is rendered from overrides/home.html.
| Section | What you find there |
|---|---|
| Getting started | Quick start, registering servers, the first workflow |
| How-to guides | Connecting clients, authentication, managing servers, Kubernetes, workflows, troubleshooting |
| Reference | Meta-tools and core tools, toolsets, configuration, custom resources, events, HTTP endpoints, CLI |
| Explanation | Architecture, aggregation, orchestration, observability, decision records |
| Operations | Installation, security |
| Contributing | Development setup, the scenario test framework |
Contributing
Issues and pull requests are welcome at github.com/giantswarm/muster.
The contributing guide explains the development setup, the
conventions the linters enforce and the behavioural test suite (muster test) that every change
is expected to extend. Commits are signed off under the DCO.
Security issues are reported through Giant Swarm's responsible disclosure process, not as public issues.
License
muster is a Giant Swarm project, released under the Apache License 2.0.
muster MCP questions
How does muster let an AI client use many MCP servers at once?
It connects to the MCP servers an organisation runs and serves all their tools through a single MCP endpoint, with tools appearing as x_<server>_<tool>.
How does muster keep tool discovery small in the context window?
Clients see thirteen meta-tools including list_tools, filter_tools, describe_tool and call_tool, and filter_tools ranks by relevance so listings stay compact.
How does muster authenticate users?
It acts as an OAuth 2.1 resource server in front of Dex, keeps one session per person and forwards the person's identity token downstream.
Can muster run on Kubernetes?
Yes, MCPServer, Workflow and WorkflowExecution are custom resources reconciled by muster, and a Helm chart ships the CRDs, RBAC and dashboards.
How do I connect an IDE to muster locally?
Run muster standalone, which starts the aggregator and a stdio bridge in one process, and point your MCP client at that command.