mcp-aggregator
Aggregates multiple MCP servers into a single Streamable HTTP endpoint, compressing tools to manage registration limits.
README
mcp-aggregator
Builds a container image that presents any number of MCP servers to a local client as a single Streamable HTTP endpoint, collapsed behind the two tools that mcp-compress-router exposes: get_tool_schema and invoke_tool.
Agent clients cap the number of registered tools — 128 for GitHub Copilot CLI — and every registered tool consumes context whether or not it is called. Routing servers through the compressor keeps them all reachable at a fixed cost of two tools.
Architecture
flowchart LR
client["MCP client"]
subgraph container["container"]
sg["supergateway<br/>PID 1"]
router["mcp-compress-router"]
sg -->|stdio| router
end
gh["github<br/>http"]
fs["filesystem<br/>stdio"]
int["internal<br/>streamable-http"]
client -->|"Streamable HTTP<br/>127.0.0.1:20000/stream"| sg
router --> gh
router --> fs
router --> int
Configuration is baked into the image at build time as build secrets, so a running container needs no host configuration beyond a state volume.
Requirements
- Node.js >= 24 on the build host.
build.cjsitself needs only 20.12 forprocess.loadEnvFile(), but the OAuth login path runsmcp-compress-routerthroughnpx, and that package requires Node 24 - Podman, or Docker via
build.engine npm installonce, for theajvschema validator
Configuration
build.json is gitignored and must be created before the first build. It is validated against build.schema.json on every run; the $schema reference enables editor completion and inline validation.
| Key | Purpose |
|---|---|
build.engine |
podman or docker |
build.name / build.tag |
Image coordinates, overridable with the IMAGE_NAME environment variable |
build.port |
Streamable HTTP port, used inside the container and for the published host binding. Optional, default 20000 |
build.args |
--build-arg values, and the source of truth for the pinned SUPERGATEWAY_VERSION and MCP_ROUTER_VERSION. The ARG defaults in the Containerfile are fallbacks |
build.podman.generate_quadlet_unit |
Emit a <name>.container systemd unit alongside the image |
mcp.servers |
Downstream servers, written verbatim into the router's mcp.json under mcpServers |
mcp.oauth |
Server names requiring an interactive login on the build host |
A server is stdio and requires command, or http / streamable-http and requires url. Unknown fields pass through to mcp.json untouched, so router features absent from the schema still work. Recognised extras include compressionLevel, allowedTools and disabledTools (picomatch globs), and enabled.
{
"$schema": "build.schema.json",
"build": {
"engine": "podman",
"name": "mcp-aggregator",
"tag": "latest",
"port": 20000,
"args": [
"MCP_ROUTER_VERSION=1.5.2",
"SUPERGATEWAY_VERSION=3.4.3"
],
"podman": {
"generate_quadlet_unit": true
}
},
"mcp": {
"servers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"description": "GitHub MCP Server",
"compressionLevel": "high",
"headers": {
"Authorization": "Bearer ${GITHUB_MCP_PAT}"
}
},
"filesystem": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/data"],
"env": {
"LOG_LEVEL": "${FS_LOG_LEVEL:-warn}"
},
"disabledTools": ["*write*", "*delete*"]
},
"internal": {
"type": "streamable-http",
"url": "${INTERNAL_MCP_URL}",
"compressionLevel": "max",
"allowedTools": ["search_*", "get_*"],
"oauth": {
"clientId": "${INTERNAL_CLIENT_ID}",
"scope": "read",
"callbackPort": 33418
}
}
},
"oauth": ["internal"]
}
}
Placeholders
Credentials belong in placeholders, not in build.json. Any ${VAR} in mcp.servers resolves against the environment, falling back to a .env file in the repository root. ${VAR:-default} is supported, and an unset placeholder without a default aborts the build. Environment values take precedence over .env.
.env and credentials.json are gitignored and excluded from the build context. They reach the image only through --secret mounts, which leave no trace in any layer.
Building
npm install # once
npm run build # or: node build.cjs
flowchart TD
bj["build.json"] --> val["validate against<br/>build.schema.json"]
val --> exp["resolve placeholders<br/>from environment and .env"]
exp --> mj["mcp.json<br/>temp file"]
cache["credentials.json<br/>repository root"] --> gate{"tokens missing for<br/>any mcp.oauth server?"}
gate -->|yes| login["interactive login<br/>on build host"]
login --> cj["credentials.json<br/>temp file"]
gate -->|no| cj
mj --> rev["CONFIG_REVISION<br/>digest of both files"]
cj --> rev
rev --> eng["engine build<br/>with both files as secrets"]
eng --> img["image"]
eng --> unit["name.container unit"]
build.cjs invokes the engine directly with no connection argument, so with Podman the image is created on whichever connection is currently default. podman system connection list shows which that is.
OAuth
The router's authorization-code flow needs a browser, which a container build cannot provide, so login runs on the build host. The resulting token store is cached in credentials.json at the repository root and later builds are non-interactive. Only servers with no stored access_token are authorized:
npm run login # or: node build.cjs --login — re-authorize every server in mcp.oauth
Deploying
With generate_quadlet_unit enabled, a successful build writes mcp-aggregator.container:
mkdir -p ~/.config/containers/systemd
cp mcp-aggregator.container ~/.config/containers/systemd/
systemctl --user daemon-reload
systemctl --user start mcp-aggregator
Quadlet honours [Install], so daemon-reload is sufficient for the unit to start at boot and there is no enable step. A rootless service that must survive logout also needs loginctl enable-linger "$USER" once.
The unit binds to 127.0.0.1 only, mounts the host CA bundle read-only, keeps router state in a named volume, and runs with RunInit=true.
Connecting a client
http://127.0.0.1:20000/stream
The path is set in the generated launcher and is not supergateway's default of /mcp. The port follows build.port.
Runtime behaviour
Configuration storage
The named volume holds router state across rebuilds, which means image content written beneath it is seeded exactly once. The two configuration files therefore live in different places:
flowchart LR
subgraph vol["named volume — persists across rebuilds"]
link["mcp.json<br/>symlink"]
cred["credentials.json<br/>rewritten on token refresh"]
end
subgraph img["image layer — replaced on every build"]
cfg["/home/mcp/config/mcp.json"]
end
link --> cfg
credentials.json persists because the router rewrites it whenever a token refreshes, and those tokens must outlive a rebuild. mcp.json is a symlink out to plain image content, so a rebuild always supplies the current server configuration.
Layer caching
Container engines do not invalidate a cached RUN when the contents of a mounted secret change. build.cjs hashes the generated mcp.json and credentials.json and passes the digest as CONFIG_REVISION, so that layer rebuilds when the configuration changes and only then.
Entrypoint
Exec-form ENTRYPOINT performs no variable substitution, so the Containerfile writes /usr/local/bin/entrypoint at build time with build.port already substituted. The script execs supergateway, which therefore remains PID 1 and receives signals directly. It exists only inside the image.
Sessions
supergateway runs in stateful mode, so one router process serves every session. In stateless mode it starts a router per session, and concurrent processes overwrite each other's refreshed OAuth tokens.
Health
supergateway registers --healthEndpoint and --cors only for SSE and WebSocket output modes, so neither is available over Streamable HTTP. The quadlet restarts the container when the process exits, and cannot detect a process that is running but unresponsive.
Build context
The Containerfile has no COPY, and .containerignore excludes everything except the Containerfile and files matching podman-build-secret-*. Both negations are load-bearing. With * alone the build cannot resolve the Containerfile, and podman-remote transports --secret sources inside the context tarball under generated names of that form, so excluding them fails the build with a server-side rename error. Docker sends secrets over the BuildKit session rather than the context, where the second negation is inert.
Files
| Path | Role |
|---|---|
build.cjs |
Build driver: validate, resolve, authorize, build, emit unit |
build.json |
Build and server configuration (gitignored) |
build.schema.json |
JSON Schema for build.json |
Containerfile |
Image definition |
.containerignore |
Deny-all with a single exception |
.env |
Placeholder values (gitignored) |
credentials.json |
Cached OAuth token store (gitignored, created on first login) |
mcp-aggregator.container |
Generated Quadlet unit (gitignored) |
Licence
MIT © Džiugas Eiva
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.