Register Streamable HTTP MCP servers, compose them into bundles, and expose each bundle through endpoints shaped for claude.ai, ChatGPT or your own scripts. Self-hosted, restart-safe, horizontally scalable.
Streamable HTTP, probed and cached
Prefixed, curated, renamed
assistant weather__forecastweather__currentnotes__searchadd_notewas notes__createcrm__lookupcrm__updatedisabledSame tools, three contracts
/assistant/mcp/assistant-gpt/mcp/assistant-ci/mcpEvery tool call passes the endpoint's auth, the bundle's curation, the client profile, the optional metering hook, and lands in the audit log. Sessions live in Valkey, so any replica serves any client.
Ten minutes from clone to a working bundle, with Docker and ports 8000, 3000 and 8080 free. The compose stack ships a demo server, bundle and endpoint, so the first call needs nothing but a key.
git clone https://github.com/mhoegger/bundlemcp.git && cd bundlemcp
cp deploy/.env.example deploy/.env
docker compose -f deploy/docker-compose.yml --env-file deploy/.env \
up -d --build --wait --wait-timeout 300KEY=bmcp_sk_devservicekey000000000000000000000000
V=2026-07-28
curl -s -N http://localhost:8000/demo/mcp \
-H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
-H "MCP-Protocol-Version: $V" -H "Mcp-Method: tools/call" \
-H "Mcp-Name: echo__reverse" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{
"name":"echo__reverse","arguments":{"text":"bundlemcp"},
"_meta":{"io.modelcontextprotocol/protocolVersion":"'"$V"'",
"io.modelcontextprotocol/clientCapabilities":{}}}}'// arrives as the last `data:` event of an SSE stream
{
"jsonrpc": "2.0",
"id": 1,
"result": { "content": [{ "type": "text", "text": "pcmeldnub" }] }
}An MCP SDK builds that envelope for you — the headers and _meta above are what a hand-rolled client has to send. See Request envelope.
Then open the admin UI at localhost:3000, register your own servers, and connect claude.ai with an OAuth endpoint. The whole path is in the Quickstart.