Installation

Bowmark's hosted MCP is at https://api.bowmark.ai/mcp. It needs no auth, so every path below is a config change and nothing else.

Want proof before you edit a config? Quickstart gets a real answer out of Bowmark with one curl and nothing installed. Come back here once you have seen it work.

The shortest path for your host

Most hosts have something better than pasting a URL into a settings pane. Take the row for yours and skip the rest of this page.

The plugin is the recommended path wherever it exists. It is the only one that wires the MCP and installs the skill that teaches an agent when to reach for it, and it shows up in the host's own plugin list so it can be turned off the same way everything else is.

ChatGPT

Go to chatgpt.com/plugins and click the + in the top right.

If the option is missing, turn on Developer mode first: Settings → Apps & Connectors → Advanced. The free tier works.

Name it Bowmark, choose Server URL, and paste:

https://api.bowmark.ai/mcp/chatgpt-app

Set No Auth, tick the confirmation box, and hit Create.

ChatGPT is the one host with a path on the end

Every other host on this page uses the bare https://api.bowmark.ai/mcp. Same server, same tools. ChatGPT needs the /chatgpt-app destination because it is the only client that demands OAuth discovery metadata before it will create a connector, and that document is published on that path alone.

On the next screen, hit Connect.

In a chat, type @ and pick Bowmark from the list. Then ask away.

Claude Code

claude mcp add bowmark --transport http https://api.bowmark.ai/mcp

Run it from any terminal. Every project picks Bowmark up automatically — there is no per-project step.

Unlike ChatGPT, you do not have to name it. The tools are in scope on every turn and Claude reaches for them on its own when a task needs the live web. Saying "use bowmark to…" just makes it deterministic when you want a specific call.

Want the skill bundled in as well? Use the Plugin path below — one install wires the MCP and the skill.

Plugin — the all-in-one path

One install pulls the skill and wires the MCP, and Bowmark shows up in the host's plugin list. This is the recommended path on Claude Code and Codex.

claude plugin marketplace add bowmark-ai/plugin
claude plugin install bowmark@bowmark-ai

codex plugin marketplace add bowmark-ai/plugin

Every other host

Same server, same URL, no auth.

[mcp_servers.bowmark]
url = "https://api.bowmark.ai/mcp"

{
  "mcpServers": {
    "bowmark": { "url": "https://api.bowmark.ai/mcp" }
  }
}

https://api.bowmark.ai/mcp

gemini extensions install https://github.com/bowmark-ai/plugin

{
  "mcpServers": {
    "bowmark": { "command": "uvx", "args": ["bowmark-mcp"] }
  }
}

{
  "mcpServers": {
    "bowmark": { "command": "npx", "args": ["@bowmark/mcp"] }
  }
}

from browser_use.mcp.client import MCPClient

bowmark = MCPClient(server_name="bowmark",
                    command="uvx", args=["bowmark-mcp"])
await bowmark.connect()
await bowmark.register_to_tools(tools)

Connect the Bowmark MCP at https://api.bowmark.ai/mcp.
No auth required. After it's connected, call the
`get_library` tool before browsing any site.

Skill only

For a host that supports the standard skill format but has no MCP client. Works with Claude Code, Claude Desktop, Cursor, Codex, and Windsurf.

npx skills add bowmark-ai/skill

Nothing installed at all

An agent with a shell or a fetch tool needs no install. Point it at the bootstrap document and it can use Bowmark immediately over plain HTTP:

https://bowmark.ai/skill.md

That page teaches the api.bowmark.ai/v1/* loop and then offers install as an upgrade. Two discovery URLs resolve to the same thing:

npx skills add https://bowmark.ai
claude plugin marketplace add https://bowmark.ai/marketplace.json

Your own code, not an agent

Everything above wires Bowmark into an agent. If you are writing the program yourself, you want the client library instead — real control flow on your machine, full autocomplete, no MCP involved.

npm i @bowmark/web

import { bowmark } from "@bowmark/web";

const { tracks } = await bowmark.music.search("aphex twin", 3);

pip install bowmark-web bowmark-web-stubs

from bowmark_web import bowmark

found = await bowmark.music.search("aphex twin", 3)

Both are MIT with zero runtime dependencies, generated from the same library manifest, released at the same version. Full detail — sessions, errors, the Node TypeScript-source gotcha, and the no-install HTTP path — is on Scripting.

Verify it works

Two checks, in order. The first proves the wiring, the second proves the wiring does something.

Wiring. Ask your agent to run this. It touches no site and costs nothing:

const library = await get_library({ query: "flights" });

An unrecognized query returns a one-line index of the whole library rather than an error, so this check can never dead-end. If you get a library back, you are wired up.

A real result. Then ask it, in plain English:

Price an RTX 5080 across every retailer you can reach, and tell me
which store is cheapest and what it costs right now.

A working agent calls get_library, then run, then answers with a store, a price and a clickable link — about ten seconds in total. If it answers with no tool call, it did not use Bowmark; it answered from memory or browsed.

In ChatGPT, type @ and pick Bowmark first — otherwise the model has no idea the tool exists and will answer from memory or browse instead. Everywhere else the tools are in scope on every turn.

Then read How it works for the two-call flow, and Scripting for the language you write against the library.

API keys

A key is additive. Anonymous requests omit the Authorization header and get a per-IP cap; keyed requests include it and draw on your plan budget.

curl -s https://api.bowmark.ai/v1/library?query=flights \
  -H "Authorization: Bearer $BOWMARK_API_KEY"

Create one at the dashboard. Keys are shown once, at creation, and stored hashed — there is no way to read one back later.

Two things a key unlocks beyond the cap:

• Runs that pause for a login. A script that hits a signed-in site returns status: "needs_user" with a single-use handoff URL. Without a key you are told to add one instead of being handed the link.
• Metered billing. 1,000 calls a month are free; past that it is $5 per 1,000.