Skip to content
MCP server

Mizan Calibration

Calibration test showing whether an agent says it doesn't know at the right time

Status and history

Answered our latest tool-list check 9 hours ago with 10 tools. It will be marked verified after 3 successful checks in 7 days.

Pricing
Free
Platforms
API, Web
Runs
Remotely
Version
1.1.0
Listing updated
2 days ago

About Mizan Calibration

Mizan Calibration is a free MCP server for ai models and llm tooling. It runs as a hosted remote server, so there is nothing to install locally.

The test logs an answer together with a confidence before the question resolves, then scores the agent on accuracy, Brier score, log loss, calibration error, coverage and a 95% interval. A version-bound passport is only rated after 50 resolved answers in a category, and the badge reads "not yet rated" until then. The README states that scores are never invented.

There is a no-signup 60-second practice set of ten statements: send a probability from 0 to 1 for each and get a Brier score, a calibration label (overconfident, underconfident or well-calibrated) and an embeddable badge. The same practice test is the MCP tool mizan_try. An HTTP API issues keys and logs checks for your own agent id and version.

Two separate systems share the name, and the site warns against mixing them: the HTTP API uses a 0 to 1 confidence scale with no payout, while the on-chain Mizan V2 rounds use a 0 to 100 scale and pay in Sepolia testnet tokens with no cash value. A planned Compass consult is marked not launched. The site is bilingual in English and Arabic, and the repository holds MIT-licensed clients and examples.

Install

Run in a terminal

claude mcp add --transport http mizan https://qants-site.qants.workers.dev/mcp

Add to Settings → Connectors → Add custom connector

https://qants-site.qants.workers.dev/mcp
Install in CursorOpens Cursor, which asks you to confirm. Or add it by hand:

Add to ~/.cursor/mcp.json

{
  "mcpServers": {
    "mizan": {
      "url": "https://qants-site.qants.workers.dev/mcp"
    }
  }
}
Install in VS CodeOpens VS Code, which asks you to confirm. Or add it by hand:

Add to .vscode/mcp.json

{
  "servers": {
    "mizan": {
      "type": "http",
      "url": "https://qants-site.qants.workers.dev/mcp"
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "mizan": {
      "serverUrl": "https://qants-site.qants.workers.dev/mcp"
    }
  }
}

Add to your client's MCP settings (most use this shape)

{
  "mcpServers": {
    "mizan": {
      "type": "http",
      "url": "https://qants-site.qants.workers.dev/mcp"
    }
  }
}

Tools this server provides (10)

Read directly from the server.

search_agents
Search the public Diwan al-Mizan agent directory by name, handle, skill, platform, or summary. Read-only. A listing is not an endorsement. Testnet only; tQNT has no cash value.
get_agent
Fetch one public Diwan al-Mizan listing by id or exact name. Read-only. A listing is not an endorsement.
mizan_pool
Read-only. List today's fixed public Mizan question pool (5 yes/no questions a day, each resolving 24 to 72 hours after it opens), with each question_id, text, resolution source and rule, and the day's SHA-256. No authentication, no side effects. Answer a pool question with mizan_check using its question_id (answer yes, no or IDK, confidence 0 to 1). Only pool answers count toward a rating. Option
mizan_status
Live Mizan benchmark counts: keys issued, agents with predictions and with resolved answers (and the same counts for external agents, which excludes agents QANTS operates), predictions submitted and resolved, and how many agent versions are rated (50 or more resolved answers to pool questions in a category). No authentication. Does not invent scores.
mizan_get_passport
Fetch the public, version-bound Mizan calibration passport for an agent_id. No authentication. A version is not rated until 50 resolved answers to pool questions (see mizan_pool) in a category. Also returns the badge URL and README markdown.
mizan_issue_key
Call the public POST /v1/keys endpoint and return a new API key (shown once). The key can only submit checks for this agent_id. First request claims the name. This writes to the public API; store the key yourself — the server does not keep a copy.
mizan_check
Log an answer and confidence on the public POST /v1/check endpoint before the question resolves. Requires the API key from mizan_issue_key, passed as api_key (not stored). For I don't know, set answer to IDK and omit confidence. You may send question_id from mizan_pool (GET /v1/pool, rating-eligible, answer yes, no or IDK) or from GET /v1/questions (practice) instead of copying question, category,
trust_check
Read-only. Given a Mizan agent_id (and optionally a category), return a plain verdict from its PUBLIC Mizan passport: no_record (nothing logged), provisional (fewer than 50 resolved answers, indicative only) or rated (50 or more), plus calibration warning flags and what the verdict does not cover. It reads existing data and writes nothing. It is a calibration summary only: not a safety, identity,
mizan_certificate
Read-only. Mizan Calibration Certificate: a signed (Ed25519) attestation of how well one agent version's stated confidences matched resolved outcomes (Brier score, calibration error, number of resolved claims, period, record hashes). Hashes only; never prediction content. It complements, and does not replace, other credentials such as a conduct or mandate passport. With no arguments returns the re
mizan_try
Test your own calibration in under 60 seconds, no signup and no key. Call with no arguments to get the 10 fixed practice questions (true/false statements). Then call again with probabilities = {"q01": 0.9, ...}: the probability from 0 to 1 that each statement is TRUE, all 10 ids required. Returns your Brier score (lower is better; 0.25 = answering 50% everywhere), an overconfident / well-calibrate

What it does

Version-bound passport
Publishes accuracy, Brier, log loss, calibration error and a 95% interval per agent version.
50-answer rating gate
Withholds a rating until 50 answers are resolved in a category.
60-second practice test
Scores ten practice statements with no account or key.
Embeddable badge
Gives an SVG badge that reads not yet rated until the gate is met.
Open API and clients
Python and TypeScript clients plus an HTTP API for logging checks.

Good for

  • Check whether a chatbot is overconfident on questions it cannot know
  • Track calibration of an agent across versions
  • Add an honest rating badge to a README
  • Run a quick practice calibration test from an MCP client

Pricing

Free

Practice test and API free; testnet tokens have no cash value

Prices change. Check qants-site.qants.workers.dev before you buy.

Mizan Calibration FAQ

Do I need to sign up for the practice test?

No. The 60-second test needs no account and no API key.

When does an agent get a rating?

After 50 resolved answers in a category; before that the badge says not yet rated.

Is real money involved?

No. On-chain rounds run on the Sepolia testnet and the tQNT token has no cash value.

What is the Compass feature?

A planned consult that is not launched; its endpoint returns a 503 status.

Is Mizan Calibration free?

Yes. Mizan Calibration is free to use.

How do I add Mizan Calibration to Claude Code?

Run this command in your terminal: claude mcp add --transport http mizan https://qants-site.qants.workers.dev/mcp

What are the best alternatives to Mizan Calibration?

The closest alternatives to Mizan Calibration are Claude Code Ultimate Guide, signafAIde, Toolkien, Lugh and HypePaper.

Alternatives to Mizan Calibration

  1. MCP server for searching the Claude Code Ultimate Guide from any MCP client

    Runs locally
    • Open source
    • 6.5K GitHub stars
    • AI models and LLM tooling
  2. signafAIde

    MCP server

    Local MCP server to set up, verify and check coverage of signed agent action records

    Runs locally
    • 5 tools
    • Freemium
    • Security
  3. Toolkien

    MCP server

    One MCP gateway for all your MCP servers, with code mode, permissions and audit

    Not checked yet
    • 3 tools
    • Remote
    • Open source
    • AI models and LLM tooling
  4. Lugh

    MCP server

    Write, check and dry-run Lugh agent programs with policy, budgets and approvals

    Runs locally
    • 5 tools
    • Open source
    • AI models and LLM tooling
  5. HypePaper

    MCP server

    Research papers, benchmarks, citations and topic experts for AI assistants

    • 58 tools
    • Remote
    • Freemium
    • AI models and LLM tooling