The official MCP Registry now lists more than 64,000 servers. Getting your own entry in is easier than it looks, especially if your server is already hosted somewhere: you do not publish code, you publish a small server.json that points at a remote endpoint. Here is the setup we have used for five versions in three weeks, with the real run times and the one failure we caused ourselves.
Disclosure: I build data tools for AI agents as Fast Data on Apify. Our registry entry points at Apify's hosted MCP server with our tools preselected; the files below are in our public data-recipes repo.
The server.json for a remote server
{
"$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
"name": "io.github.ichkifokkus/fast-data",
"title": "Fast Data Actors (via Apify MCP)",
"description": "German/EU e-invoices, IBAN/VAT/BaFin checks, tenders, hiring signals and more via Apify",
"repository": { "url": "https://github.com/ichkifokkus/data-recipes", "source": "github", "subfolder": "mcp" },
"version": "1.1.3",
"remotes": [
{
"type": "streamable-http",
"url": "https://mcp.apify.com/?tools=fast_data_asap/iban-validator-bank-lookup,fast_data_asap/eu-vat-number-validator",
"headers": [
{
"name": "Authorization",
"description": "Apify API token, e.g. 'Bearer <token>'. Optional for clients that support OAuth.",
"isRequired": false,
"isSecret": true
}
]
}
]
}
(The real file lists 18 tools in the URL; shortened here.)
| Field | What to know |
|---|---|
name |
Namespace and name. io.github.<user>/… is what GitHub-based auth lets you publish; a custom domain namespace needs DNS or HTTP verification instead. |
version |
Must change on every publish. The registry keeps every version and marks the newest as latest. |
remotes[].type |
streamable-http for a hosted endpoint. Local servers use packages (npm, PyPI, …) instead. |
headers[] |
Declares what a client must send. isSecret: true tells clients to treat it as a credential and never log it. |
description |
Short. It is what people see in registry listings and client UIs. |
The workflow: GitHub OIDC, no stored secrets
The registry accepts a GitHub Actions OIDC token as proof that the workflow runs in the repository that owns the io.github.<user> namespace. No API key goes into the repo or the secrets store.
name: Publish to MCP Registry
on:
push:
branches: [main]
paths: [server.json]
workflow_dispatch:
jobs:
publish:
runs-on: ubuntu-latest
permissions:
id-token: write # required for mcp-publisher login github-oidc
contents: read
env:
MCP_PUBLISHER_VERSION: v1.8.8
steps:
- uses: actions/checkout@v5
- uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
- name: Download and verify mcp-publisher
run: |
set -euo pipefail
archive="mcp-publisher_linux_amd64.tar.gz"
base="https://github.com/modelcontextprotocol/registry/releases/download/${MCP_PUBLISHER_VERSION}"
curl -fsSL -o "$archive" "$base/$archive"
curl -fsSL -o "$archive.sigstore.json" "$base/$archive.sigstore.json"
cosign verify-blob --bundle "$archive.sigstore.json" \
--certificate-identity "https://github.com/modelcontextprotocol/registry/.github/workflows/release.yml@refs/tags/${MCP_PUBLISHER_VERSION}" \
--certificate-oidc-issuer "https://token.actions.githubusercontent.com" "$archive"
tar xzf "$archive" mcp-publisher
- run: ./mcp-publisher validate
- run: ./mcp-publisher login github-oidc
- run: ./mcp-publisher publish
Three choices worth copying:
-
Trigger on
paths: [server.json]. A publish happens exactly when the entry changes, never on unrelated commits. -
Pin the publisher version and verify its signature. Downloading
releases/latestand running it with an OIDC token in scope is a supply-chain risk; cosign checks that the binary was built by the registry's own release workflow. -
validatebeforelogin. Schema errors fail fast without touching the registry.
What it looked like in practice
| Version | Published (UTC) | What changed |
|---|---|---|
| 1.0.0 | 22 Oct, 20:29 | first entry |
| 1.1.0 | 8 Nov, 13:46 | finance and sales tools added |
| 1.1.1 | 8 Nov, 20:22 | EU tenders tool |
| 1.1.2 | 10 Nov, 09:31 | company check tool |
| 1.1.3 | 11 Nov, 09:29 | exchange-rate converter |
Each successful run took 11 to 14 seconds from start to finish. You can check your own entry, including every version, with one public call:
curl -s "https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.ichkifokkus" \
| python3 -c "import json,sys; [print(s['server']['version'], s['_meta']['io.modelcontextprotocol.registry/official']['isLatest']) for s in json.load(sys.stdin)['servers']]"
The error you will cause yourself
On 8 November I re-ran the workflow manually twice, 13 minutes after 1.1.0 had gone through. Both runs failed. The server.json on that commit still said "version": "1.1.0", and the registry does not accept a version that already exists. Nothing was broken; the entry was already live.
The fix is procedural: bump version in the same commit as the change. Since the workflow triggers on server.json, a forgotten bump means the publish step fails, and a manual re-run of the same commit cannot help.
What a registry entry does not do
-
It does not host anything. The URL in
remotesmust already work; the registry only stores metadata. - It is not a quality signal. With roughly 670 new servers a day, a listing alone gets little attention. Clear descriptions and a useful tool set matter more.
- It does not handle auth for you. Declaring a header tells clients what to send; your endpoint still has to check it.
Try it
Copy server.json and .github/workflows/publish-mcp.yml from our data-recipes repo, change the namespace, URL and description, and push. The entry shows up in the registry API within seconds of a green run.
Question: for those publishing local servers via packages: do you publish a registry version on every package release, or only when the tool surface changes? We bump on every tool change, but I am not sure that scales.
Top comments (0)