Claude Fable 5.2 MCP: setup and a real run with costs

To use Claude Fable 5.2 with MCP, add an MCP server such as looot (https://api.looot.ai/mcp) to Claude Code, claude.ai or the Claude API, then pick the model claude-fable-5-2. In our run on 2026-11-09, Fable 5.2 made 69 looot calls to find 10 leads, and the average cost per call ran from a fraction of a cent for a page of Google results to about two cents for a people search.

Anthropic released Fable 5.2 the same day. It is the model Anthropic points at long, many-step agent work, and a lead hunt with dozens of lookups is that kind of job. Below are the three setups, the receipts from that run, and when Opus 5.5 is the better pick. The full walkthrough, with the exact prompt and every lead, is in the docs guide. The model choice gets its own page in Fable 5.2 vs Opus 5.5 for agent data work.

How do I add an MCP server to Claude Code with Fable 5.2?

Add the server once, sign in on looot.ai in the browser, then start Claude Code on Fable 5.2. Three commands. The server is added for every project, so you do it once per machine.

# 1. Add looot for every project
claude mcp add --scope user --transport http looot https://api.looot.ai/mcp

# 2. Sign in: pick the organization to bill and click Connect
claude mcp login looot

# 3. Start a session on Fable 5.2
claude --model claude-fable-5-2

Type /mcp and check that looot shows Connected. Then ask for your balance. That call is free and proves the connection reaches your organization. On a server with no browser, pass an agent token as an Authorization header instead, as the Claude Code connect page shows.

How do I use Fable 5.2 with MCP in claude.ai?

One custom connector covers claude.ai, Claude Desktop and the Claude mobile apps. You sign in on looot.ai, so there is no token to paste.

1. Open Settings, then Connectors. Click Add, then Add custom connector.
2. Name it looot and set the MCP server URL to https://api.looot.ai/mcp. Leave Advanced settings empty.
3. Click Continue, sign in on looot.ai, pick the organization and click Connect.
4. In a chat, open the + menu, then Connectors, and turn looot on. Pick Claude Fable 5.2 in the model menu.

On Team and Enterprise plans, an Owner adds the connector once for the organization and each member clicks Connect. The claude.ai connect page has screenshots.

How do I call claude-fable-5-2 with an MCP server from the Claude API?

Use the MCP connector on the Messages API. It takes two parts: an mcp_servers entry with the URL, and an mcp_toolset tool with the same name, sent with the beta flag mcp-client-2025-11-20. Anthropic's cloud opens the connection, so browser sign-in does not apply. Create an agent token in your looot settings, keep it in an environment variable, and pass it as authorization_token.

import os
import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    betas=["mcp-client-2025-11-20"],
    mcp_servers=[
        {
            "type": "url",
            "url": "https://api.looot.ai/mcp",
            "name": "looot",
            "authorization_token": os.environ["LOOOT_TOKEN"],
        }
    ],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "looot"}],
    messages=[{"role": "user", "content": "Use looot: check my balance, then find and verify the work email of the CEO of example.com."}],
)

for block in response.content:
    if block.type == "text":
        print(block.text)

Check stop_reason before you read the answer. Fable 5.2 thinks on every request, so a hard task can run for minutes: give the client a long timeout, or stream. The headless setup page covers tokens for CI and scripts.

What did a real Claude Fable 5.2 MCP run cost per call?

We gave Claude Code on claude-fable-5-2 one task and no tools except looot: find 10 leads at firms that sell open-source business software, a founder or CEO at each, that person's work email checked, and the firm enriched. It could spend at most one dollar of looot credit. It ran 69 jobs over 74 turns in about 15 minutes and stopped well under the cap. Here is what each job cost.

The people search carried most of the bill. The first provider in line, AI Ark, missed all 18 searches and still billed for each one, which was about two thirds of the data spend. LeadMagic then found the person on 17 of them. The email finder shows the other kind of miss: when Kitt AI found nothing, it charged nothing, and Filtrat answered. Read route.attempts on your own runs to see the same split.

Of the 10 leads, 5 ended with a verified work email and 5 sit on catch-all domains, where the verifier cannot confirm the mailbox. Treat those 5 as unconfirmed. Here is one lead from start to finish, with its run ids:

The data side is only part of what the session cost. The model tokens are the other part, and with Fable 5.2 they were the larger one by far.

Most of the token bill is Claude Code writing the growing conversation to the prompt cache. The looot side would be the same with any other model, since a lookup costs the same whichever model asks for it.

Should I run my MCP agent on Fable 5.2 or Opus 5.5?

Start with Opus 5.5 unless you have a reason not to. That is Anthropic's own advice: Opus 5.5 for most workloads, Fable 5.2 for demanding reasoning and long-horizon agent work, or when Opus 5.5 at higher effort still falls short (models overview, read 2026-11-09).

Fable 5.2 lists at $8 per million input tokens and $40 per million output tokens, twice the $4 and $20 of Opus 5.5 (Anthropic pricing, read 2026-11-09). Both have a 1M token context window and 128K max output. Fable 5.2 defaults to high effort, Opus 5.5 to medium.

We did not run the same task on Opus 5.5, so the run above shows what one Fable 5.2 session cost, not which model is better at it. The comparison guide works through when the extra token cost pays off. For model prices side by side, see AI model prices.

What looot does and does not do

• Prepaid, per call. A workspace starts at zero and spends credit you top up, one call at a time. The pricing page explains how a charge is worked out.
• A price before you spend. search and inspect are free and show each endpoint's price and the most a run can cost, so the agent can check before it runs.
• Fallback on a miss. With fallback, a run that misses tries the next provider for the same job, up to the maxCostUsd you set. A provider that ran and found nothing can still bill, as AI Ark did above.
• Reads data, does not act as you. looot does not post on social media, log in to sites as you, buy things or read private accounts.

If you want the same server in other clients, Claude MCP servers for data work lists what else Claude can reach, and one prepaid balance for an MCP server explains the billing model.