# Muse Wonder MCP: give Meta's model live web data

Muse Wonder and MCP: Meta's API has no MCP field, so you add MCP servers in Muse Code, Claude Code or Codex. The config, the tool choice and the limits.

## Summary

Muse Wonder can use MCP tools, but not through Meta's API on its own. The API has no MCP field: MCP lives in the client. Muse Code takes MCP servers in its settings file, and Claude Code or Codex running `muse-wonder-1.0` keep their own MCP servers. Your own app can also run any tool itself.

That split is the most common confusion since Meta released Muse Wonder on October 12, 2026. People ask whether the model can search the web, read a page or look up a company, and the answer depends on where it runs. This post gives the config for each place, explains when the built-in web search is enough, and lists what a data tool adds. At the end you will have one MCP server feeding Muse Wonder live data in the client you already use.

## Does Muse Wonder support MCP?

Yes, through its clients. Meta's tool-calling page says "the model does not execute developer-defined tools itself. Your app handles execution". The client is the app. Muse Code reads an `mcp_servers` block, and Meta's [coding agents page](https://dev.meta.ai/docs/coding-agents) marks MCP server tools as supported in Claude Code and Codex on Muse Wonder.

Meta's Messages endpoint rejects unknown top-level fields, so there is no server-side MCP connector to switch on. Pick one of the three options below.

## Option A: Muse Code with an mcp_servers entry

Muse Code takes two MCP transports: `stdio` (a command it starts) and `streamable_http` (a URL with headers). looot's MCP server is a streamable HTTP server at `https://api.looot.ai/mcp`. Add it to the `mcp_servers` block of your Muse Code [settings file](https://dev.meta.ai/docs/muse-code/extending):

Three details from Meta's docs:

- `${LOOOT_TOKEN}` is read from your environment, so the token stays out of the file. `npm install -g looot && looot init` prints the export line for it.
- `mode` defaults to `required`, which aborts the whole run when the server is down. `optional` skips it with a warning.
- Muse Code can also sign in to a remote server with OAuth: drop the `headers` line and run `muse mcp login looot`.

Run `/mcp` in a Muse Code session to see the connected servers and their tools. We set looot up this way inside Muse Code on 2026-10-13, and it connected like any other MCP client: add the server, sign in, and the tools appear. Muse Wonder Full is in Muse Code only, on the Power Usage and Maximum Usage plans, so this is also the only way to give Full your tools.

## Option B: Claude Code or Codex running muse-wonder-1.0

Meta's coding agents page shows Claude Code on Meta's API: `ANTHROPIC_BASE_URL` set to `https://api.meta.ai`, your Meta key in `ANTHROPIC_AUTH_TOKEN`, and `ANTHROPIC_MODEL` set to `muse-wonder-1.0`. Set `ENABLE_TOOL_SEARCH=true` as well, because Claude Code turns off MCP tool search for hosts other than Anthropic.

Then add looot the usual way:

For Codex, use the model config on Meta's page, then `codex mcp add looot --url https://api.looot.ai/mcp` and `codex mcp login looot`. Your MCP servers do not change when the model does. If you are still choosing the model, [Muse Wonder vs Claude Opus 5.5 vs Gemini 4 Argon](https://looot.ai/blog/muse-wonder-vs-claude-opus-5-5-vs-gemini-4-argon) compares the three on price and access.

## Option C: your own app on Meta's API

If you call Meta's Responses or Messages API from your own code, define function tools and run them yourself. Meta's API supports parallel tool calls and structured output. Inside your tool handler, call looot over HTTP: one key, one balance, and the run tells you what it cost. [LLM tool calling with live APIs](https://looot.ai/blog/llm-tool-calling-with-live-apis) walks through the loop.

## Built-in web search or a data tool?

Use Meta's built-in `web_search` when you want a grounded answer with citations and nothing more. Meta charges $2.50 per 1,000 search queries on top of tokens ([Meta pricing](https://dev.meta.ai/docs/pricing-rate-limits), read 2026-10-13).

Use a data tool when the agent needs rows it can work with:

| The agent needs | Built-in web_search | A data tool through MCP |
|---|---|---|
| A cited answer to a question | yes | yes, from search results |
| The full text of one page | no | URL to markdown |
| Raw search results to rank or filter | no | web search API |
| Posts on X or Reddit about a topic | no | X search, Reddit search |
| A company record or a work email | no | enrich a company |

Price per call for the data tools above. Prices read on 2026-10-14 from the live looot catalog. They change when a provider changes its price.

| Option | Cheapest provider | Price per call |
|---|---|---|
| Search the web | Firecrawl | $0.0010 per call |
| Read a page as markdown | Jina AI | $0.0002 per call |
| Enrich a company | The Companies API | $0.0019 per call |
| Find a work email | Tomba | $0.0089 per call |

## Three things to try with live data

Each one is a job id the agent passes to looot's `run` tool. The agent can `inspect` a job first to read its price, which is free.

1. What people say about a launch: `job:x.post.search`, `job:reddit.post.search` and `job:news.search`, then Muse Wonder sums it up with links. The research for this post started that way.
2. Research a page: `job:web.search` to find it, `job:web.scrape.markdown` to read it. Long pages fit in Muse Wonder's 1,048,576-token context window.
3. Account research: `job:company.enrich` for the company, `job:people.email.find` for a contact.

## What it cannot do

- Meta's API does not connect to MCP servers by itself. Something on your side runs the tools.
- looot reads data and calls provider APIs. It does not post on X or Reddit, log in to sites as you or read private accounts.
- Every data call is paid from your looot balance. A miss is free unless the provider bills for one, and the run says which.
- MCP inside the Muse consumer app is not covered here. This post is about Muse Code, coding agents and your own code.
- Muse Wonder is in preview. Meta can change the API, the price or the client config before general release.
- Muse Code tools are not sandboxed: Meta's docs say an MCP server runs outside the sandbox, so connect only servers you trust.

Last checked 2026-10-13. Author: the looot team.

## Questions

### Does Muse Wonder support MCP?

Yes, through the client it runs in. Muse Code reads an `mcp_servers` block, and Claude Code and Codex keep their MCP servers when they run `muse-wonder-1.0`. Meta's API itself has no MCP field.

### How do I add an MCP server to Muse Code?

Add an entry to the `mcp_servers` block of your Muse Code settings file, with `transport` set to `stdio` or `streamable_http`. For a remote server, give the `url` and either a `headers` entry with a token or an OAuth sign-in with `muse mcp login`.

### Can I use Muse Wonder in Claude Code with my MCP servers?

Yes. Set Meta's environment variables from its coding agents page and add `ENABLE_TOOL_SEARCH=true`, then your MCP servers load as before. The model changes, the tools do not.

### Does Muse Wonder have built-in web search?

Yes, a `web_search` tool on Meta's API. It returns grounded answers and costs $2.50 per 1,000 search queries on top of tokens ([Meta pricing](https://dev.meta.ai/docs/pricing-rate-limits), read 2026-10-13).

### How much does web search cost with Muse Wonder?

Meta's built-in search grounding is $2.50 per 1,000 queries ([Meta pricing](https://dev.meta.ai/docs/pricing-rate-limits), read 2026-10-13). A data tool through MCP has its own price per call, shown in the live table above.

### Can Muse Wonder scrape a website?

Not by itself. It reads what a tool hands it: a page as markdown from `job:web.scrape.markdown`, or what it sees through computer use. The tool does the fetching.

### Can Muse Wonder Full use MCP tools?

Yes, in Muse Code. Muse Wonder Full is available only there, on the Power Usage and Maximum Usage plans, and Muse Code's `mcp_servers` block works the same for both variants.

### Do I need an MCP server, or can I call APIs directly?

You can call APIs directly from your own tool handlers on Meta's API. An MCP server saves the work in a client like Muse Code or Claude Code, where you cannot write the tool loop yourself.

## For agents

### Code (json)

```json
{
  "mcp_servers": {
    "looot": {
      "transport": "streamable_http",
      "url": "https://api.looot.ai/mcp",
      "headers": { "Authorization": "Bearer ${LOOOT_TOKEN}" },
      "mode": "optional"
    }
  }
}
```

### Code (bash)

```bash
claude mcp add --scope user --transport http looot https://api.looot.ai/mcp
claude mcp login looot
```

### Give Muse Wonder live data on a launch

```text
Use looot to find out what people say about a product launch, from the agent running on Muse Wonder.
1. Set up https://looot.ai/skill.md.
2. Inspect job x.post.search, job reddit.post.search and job news.search, and tell me the price of each before you run anything.
3. Run each one for "Muse Wonder" from the last 7 days.
4. Sum up the three main reactions with a link for each, then print every run's outcome and what it cost.
```