OP3 MCP Server & CLI README

OP3 MCP

OP3 sits in front of the audio file, so it counts every download whatever app made it.

That is why its numbers differ from Apple's or Spotify's, which each report only their own listeners.

22 tools, including unique listeners, retention cohorts and episode benchmark curves that OP3's own API does not expose.

Built and maintained by Navid Moazzez. Built on Slipway, which turns one definition of each tool into the MCP server and the CLI.

Two ways to use it

Command line

op3-cli runs every tool as a command. Agents that run commands, like Claude Code, Codex and OpenCode, use it on their own, and you can type the same commands in a terminal, a script or a cron job:

npm install -g @thenavidm/op3-mcp-cli
op3-cli                                   # every command
op3-cli op3-global-app-share --agent
op3-cli op3-get-show --identifier https://feeds.example.com/show.xml
op3-cli op3-show-downloads --help
op3-cli which listener retention          # find the command for a task

Every command reads OP3's public data, so nothing asks for --confirm. --agent is JSON on one line for scripts and agents, and --select a,b.c keeps only the fields you name.

op3-cli schema <command> prints the exact JSON Schema an MCP client receives for that tool.

MCP server, for AI agents

op3-mcp is what Claude Code, Claude Desktop, Cursor and the rest launch. You never run it by hand:

claude mcp add op3 -- npx -y @thenavidm/op3-mcp-cli

In Claude Desktop, the .mcpb extension installs on a double click. Section 4 has every other client.

Contents

1. What you can ask it
2. Quick install
3. Setup
4. Connect your client
5. Check it worked
6. What it costs to have connected
7. Tools
8. Reading the numbers
9. Your data
10. Troubleshooting

1. What you can ask it 💬

• How many downloads did my podcast get this month?
• How many actual people is that, not downloads?
• Are my listeners coming back, or is it a new audience every month?
• Is this week's episode tracking ahead or behind my usual?
• Where in the world is my audience, down to the state?
• Which podcast apps do my listeners use, and is that unusual?
• Compare my show against these three others.
• What day and hour do most downloads happen?
• Which of my episodes share the same listeners?
• OP3 shows nothing for my feed. What is wrong?

The one that is impossible without this server is the second. Every podcast analytics dashboard reports downloads. A download is an app fetching a file, not a person, and one listener whose app re-requests across three days counts three times. OP3's raw rows carry a privacy-preserving per-listener hash, so unique listeners, returning listeners and retention can be computed from them. None of OP3's own aggregated endpoints expose that, and this server does.

2. Quick install ⚡

Node 22 or newer. Nothing else.

npx -y @thenavidm/op3-mcp-cli@latest --version

That is the whole install. npx fetches it on demand, so there is nothing to update later: with @latest, a new version reaches you the next time your client starts the server.

To build from source instead:

git clone https://github.com/thenavidm/op3-mcp-cli.git
cd op3-mcp
npm install
npm run build
npm test
node dist/index.js --version

3. Setup 🔑

You can skip this. The server works with no credential at all, because OP3 publishes a shared preview token and that is what it falls back to. It is rate limited and OP3 can withdraw it, so get your own before relying on it.

Have an agent do it

The agent cannot sign in to OP3 for you. What it can do is walk you through it and wire up the config.

Paste this into Claude Code, Cursor, or any agent with terminal access:

Help me set up the OP3 MCP server.

1. Open https://op3.dev/api/keys and tell me what to click to create an
   API key and its bearer token.
2. Stop and wait for me to paste the token back.
3. Add the server to my MCP client config with that token as OP3_TOKEN.
4. Run the doctor command and tell me whether it worked.

Or do it yourself

1. Go to op3.dev/api/keys.
2. Create an API key and copy its bearer token.
3. Set it as OP3_TOKEN in your client config, as in section 4.

You do not need to own a podcast. The token reads public OP3 data, which covers every show that has the OP3 prefix on its feed.

To revoke

Delete the key at op3.dev/api/keys. It stops working immediately.

If you want your own show in here

Your podcast needs the OP3 prefix on its episode audio URLs. That is a change to your feed, not to this server, and it is documented at op3.dev/setup. Until a listener downloads an episode through the prefix, OP3 has no data for your show and neither does this.

4. Connect your client 🔌

The long version, every step with what to do when one fails, is in INSTALL.md.

OP3_TOKEN is optional in every block below. Leave it out to use OP3's preview token.

Claude Code

claude mcp add op3 \
  -e OP3_TOKEN=your-token \
  -- npx -y @thenavidm/op3-mcp-cli@latest

Add --scope user to make it available in every project rather than just this one.

Claude Desktop

The short way: download the .mcpb extension from the latest release and double-click it. It carries its own dependencies, so there is no config file to edit and nothing to install first. It asks for an OP3 token, which is optional.

The long way, if you would rather edit the config yourself:

{
  "mcpServers": {
    "op3": {
      "command": "npx",
      "args": ["-y", "@thenavidm/op3-mcp-cli@latest"],
      "env": { "OP3_TOKEN": "your-token" }
    }
  }
}

Tip Claude Desktop does not inherit your shell PATH. If npx is not found, use the absolute path from which npx.

Quit Claude Desktop completely and reopen it.

claude.ai on the web

claude.ai runs a connector from Anthropic's cloud rather than your machine, so it cannot launch a local command. It needs a public HTTPS URL.

Run the server over HTTP:

npx -y @thenavidm/op3-mcp-cli@latest --http --port 8000

Host it somewhere with a public HTTPS URL, then in claude.ai go to Customize, Connectors, +, Add custom connector, paste the URL and click Add.

On Team and Enterprise plans an owner adds it first under Organization settings, Connectors, then each member enables it under Customize, Connectors.

It binds to 127.0.0.1 by default, which is all a tunnel or a proxy on the same machine needs. To accept connections from outside the machine, set OP3_HTTP_HOST=0.0.0.0 together with OP3_HTTP_TOKEN, a bearer token every request must send: without one, the server refuses to listen beyond this machine.

Cursor

.cursor/mcp.json, same JSON shape as Claude Desktop, key mcpServers.

Windsurf

~/.codeium/windsurf/mcp_config.json, key mcpServers.

VS Code

.vscode/mcp.json. The key is servers, not mcpServers, and each entry takes "type": "stdio".

{
  "servers": {
    "op3": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@thenavidm/op3-mcp-cli@latest"],
      "env": { "OP3_TOKEN": "your-token" }
    }
  }
}

Codex CLI

~/.codex/config.toml:

[mcp_servers.op3]
command = "npx"
args = ["-y", "@thenavidm/op3-mcp-cli@latest"]

[mcp_servers.op3.env]
OP3_TOKEN = "your-token"

Gemini CLI

~/.gemini/settings.json, key mcpServers.

Everything else

Any stdio MCP client takes the same three things: the command npx, the args, and the env block.

Docker

docker build -t op3-mcp .
docker run --rm -i -e OP3_TOKEN=your-token op3-mcp

5. Check it worked 🩺

npx -y @thenavidm/op3-mcp-cli@latest doctor

It tests the token against the cheapest OP3 endpoint, tests the raw query endpoints separately because they fail for their own reasons, and prints the settings in force.

6. What it costs to have connected

Both surfaces are the same program with the same 22 tools. The difference is when the model pays for them. Measured in Claude Code:

Claude Code's tool search is on by default: it sends only the tool names and the server instructions, and loads a tool's full definition when the model reaches for it. An app that loads every tool up front pays the first line on every message, whether OP3 comes up or not. With the skill added, Claude Code also lists its one-line description, about 190 tokens.

To spend less, turn the server off when you are not using it, which in Claude Code is the /mcp panel. Or install the CLI and add the server on the days it earns its place.

Measured on 2026-10-05 against 1.2.2, with Claude Code 2.1.286 on Claude Opus 5.5 (one short prompt with and without the server connected, once with ENABLE_TOOL_SEARCH=false and once with the default, the difference read from the API's own usage figures; SKILL.md the same way) and Codex 0.159.3 on gpt-6.1-sol:

The task was "find the command that compares listener retention between two periods, and the flags it requires". Every tool loaded costs less because each tool no longer repeats $schema, additionalProperties and an execution block. Over the CLI, every 1.2.2 run read the general help, the command list and the command's help, three requests that each carry the conversation so far; which answers with the command's help when one command fits well ahead of the rest, so 4 of five 2.0.0 runs needed two. Over MCP, Codex answered from the tool list without calling a tool, and read about 40,000 input tokens in one run of five on each version and about 78,000 in the rest. SKILL.md costs 40 more because it lists which and exit code 1. Other apps and models count tokens a little differently, and tool-list characters divided by four are not API usage.

7. Tools 🛠️

Every tool is a read. Nothing here changes anything.

Finding a show

Downloads, from OP3's rolled-up data

Fast. Prefer these whenever they answer the question.

Audience, from the raw rows

The tools OP3's own API cannot give you.

Where and how

Over time

Discovery and setup

8. Reading the numbers 📊

The part worth more than the upstream API docs. All of it was found by probing the live API, because OP3's OpenAPI document declares no response schemas.

A download is not a person. It is an app fetching a file. Podcast analytics across the industry quotes downloads, and a show with a loyal audience whose apps re-request looks larger than a show with more actual listeners. op3_audience_summary gives both numbers and the ratio.

Episodes are not comparable on totals. An older episode has had longer to accumulate downloads. op3_episode_curve compares at equal age against the show's own median, and excludes episodes younger than the horizon from that median so a three-day-old episode does not drag a thirty-day comparison down.

App share on its own says almost nothing. Nearly every show's biggest app is Apple Podcasts, because Apple is around 38% of all podcast listening. A show at 40% Apple is under-indexed. op3_benchmark_apps divides by the global share so the number means something.

bots does more than add bots. With it off, OP3 applies its published download calculation, which deduplicates repeat requests. With it on you get raw request rows. On one probe the same window returned 213 against 345, and only 79 of the difference were bot rows.

Rolled-up figures lag about a day. They carry an asof date. That is why they will not match a live dashboard exactly.

An empty episode list is normal for a small show. OP3's daily rollup has not covered it yet. The show-level total may still be there.

Metro codes are US-only. They are a US broadcast concept. Use level=region for a worldwide breakdown.

Raw download rows come back oldest first. OP3 offers no way to reverse that on this endpoint, so a limit takes the earliest rows in the window. Narrow the window to see recent activity rather than raising the limit.

Some shows cannot be filtered by episode. OP3 derives the episode id from the episode audio URL, so a host that regenerates those URLs leaves historical rows carrying ids that no longer appear in the feed. Checked across three shows, two matched exactly and one had no overlap at all. op3_episode_curve detects this and says so rather than returning silent zeros.

The raw endpoints are scans, not indexes. Two rows took 2790ms when probed. Cost grows with the window. The server pages with a continuation token, caps what it will pull, and labels any result that was cut short, because a truncated result presented as complete makes every rate computed from it wrong.

9. Your data 🔐

There is no backend. The server runs on your machine, talks to op3.dev, and stores nothing on disk.

Responses are cached in memory for five minutes so an agent asking several questions about one show does not pay for the same scan repeatedly. That cache dies with the process. Set OP3_CACHE_TTL_MS=0 to disable it.

Per-listener identifiers never leave the server. OP3's raw row carries audienceId and hashedIpAddress. Both are aggregated over inside the process and stripped from anything returned. Every audience figure is a count, a rate or a distribution.

This is deliberate. OP3 exists to be a privacy-preserving analytics service, and a wrapper that streamed stable per-listener keys into a model context would undo that. op3_query_downloads can emit a shortened, non-reversible label if you need to tell rows apart by listener, and it is off by default.

Third-party text. Show and episode titles come from arbitrary RSS feeds. Anyone can publish a podcast, so that text is attacker-controlled and reaches the model inside a tool result. The server neutralises fence-breaking and tells the model in its instructions to treat it as data. That raises the cost of an injection rather than removing it. The real reason the blast radius is small is that this server is read-only and reaches nothing but OP3.

Settings

10. Troubleshooting 🔧

Run doctor first. It names the problem in one command.

npx -y @thenavidm/op3-mcp-cli@latest doctor

FAQ ❓

Questions

Run into a problem or have a question? Open an issue and I will help.

About the author 👋

Navid Moazzez is a leading AI business strategist, and the host of the AI Creator Summit, watched by 100,000+ creators. He helps creators and founders master AI and build their own AI Operating System (AI OS) to automate their business and life. He creates useful free tools, MCP servers and CLIs that creators and founders can use in their own workflows.

Links

• Personal website: navid.me
• Navid Media: navid.media
• YouTube: @thenavidm and @thenavidai
• X: @thenavidm
• Instagram: @thenavidm
• LinkedIn: thenavidm

If this is useful, star the repo and come say hi on X.

Dependencies

Nothing else. The OP3 client, the pagination, the aggregation and the time handling are all built in.

License

MIT. Free to use, modify, and share.

Not affiliated with, endorsed by, or connected to the Open Podcast Prefix Project.

© 2026 Navid Media. Made with ❤️ by Navid Moazzez.