# mizan-ledger

> mizan-ledger is an open-source MCP server for agent benchmarks. Local calibration ledger that scores an agent's declared confidence against real outcomes. It runs locally on your machine and is started by your AI client.

- **Kind:** MCP server
- **Pricing:** Open source
- **Platforms:** CLI, API, self-hosted, Linux, macOS, Windows
- **Website:** https://qants-site.qants.workers.dev/library
- **Repository:** https://github.com/qants-labs/mizan-ledger
- **MCP Registry name:** io.github.qants-labs/mizan-ledger
- **Categories:** [agent benchmarks](https://ai-powered.com/categories/agent-benchmarks)
- **Listing updated:** 2026-11-19

## About

The README describes mizan-ledger as a way to record how confident an agent is before a task, store only a salted SHA-256 commitment of that declaration, and later score it against the result of the agent's own tests or checks. Scoring covers Brier score, ECE, log loss, accuracy and coverage of declared, resolved and pending records.

The library runs entirely on the local machine with no network access or telemetry, and the core has no required dependencies. Public receipt and resolution files can be checked offline with a verify command that walks the hash chains, while openings are kept in a private directory that should stay out of version control.

It suits developers who want to measure whether an agent's stated certainty matches how often it is right. The project is labelled early alpha, and the README is explicit that scores measure calibration rather than whether an agent is good, safe or secure.

## Features

- **Hash-committed confidence declarations:** Each declared probability is stored as a salted SHA-256 commitment before the outcome is known.
- **Calibration scoring:** Computes Brier score, expected calibration error, log loss and accuracy with confidence intervals.
- **Offline chain verification:** A verify command checks hash chains, openings and scores without any network access.
- **Pytest plugin:** Marks tests with a confidence value and records pass or fail as the outcome.
- **Optional local MCP server:** A stdio server exposes commit, resolve, calibration and verify tools to MCP clients on the same machine.
- **Coverage reporting:** Counts declared, resolved and pending records so unresolved answers cannot be silently dropped.

## Good for

- Measuring whether a coding agent's stated confidence matches its test pass rate
- Tracking calibration of an agent across repeated tasks in a pytest suite
- Checking a local agent's confidence history offline with a tamper-evident chain
- Comparing calibration of two agent versions over a set of resolved tasks

## Install

### Claude Code (Terminal)

```sh
claude mcp add mizan-ledger -- uvx mizan-ledger
```

### Claude Desktop (claude_desktop_config.json)

```json
{
  "mcpServers": {
    "mizan-ledger": {
      "command": "uvx",
      "args": [
        "mizan-ledger"
      ]
    }
  }
}
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "mizan-ledger": {
      "command": "uvx",
      "args": [
        "mizan-ledger"
      ]
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "mizan-ledger": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "mizan-ledger"
      ]
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "mizan-ledger": {
      "command": "uvx",
      "args": [
        "mizan-ledger"
      ]
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "mizan-ledger": {
      "command": "uvx",
      "args": [
        "mizan-ledger"
      ]
    }
  }
}
```

## Tools

- `commit_confidence`: Declare confidence before the task.
- `resolve_outcome`: Record the real outcome of a declared task.
- `get_calibration`: Read-only calibration scores.
- `verify_ledger`: Read-only offline check of the ledger.

## FAQ

### What license is mizan-ledger under?

The README states it is MIT licensed.

### Does the library send any data over the network?

The README says the core library has no network access, no telemetry and no dependencies.

### Can the library verify that outcomes are honest?

No. The README says it cannot tell whether an outcome supplied by the caller is honest, and recommends an outcome source the agent does not control.

### How is it installed?

With pip install mizan-ledger on Python 3.9 or later. The optional MCP server requires pip install 'mizan-ledger[mcp]' and Python 3.10 or later.

### Is mizan-ledger free?

Yes. mizan-ledger is open source, so you can use and self-host it for free.

### How do I add mizan-ledger to Claude Code?

Run this command in your terminal: claude mcp add mizan-ledger -- uvx mizan-ledger

### What are the best alternatives to mizan-ledger?

The closest alternatives to mizan-ledger are Mizan Calibration and Agent Reliability.

## Where this comes from

- Description: written from the project's README, the project's website and its MCP Registry entry.
- Install steps: from the official MCP Registry entry.

## Alternatives

- [Mizan Calibration](https://ai-powered.com/mcp/mizan-calibration): Mizan Calibration is a free MCP server for ai models and llm tooling.
- [Agent Reliability](https://ai-powered.com/mcp/agent-reliability): Agent Reliability is an MCP server for agent benchmarks.

---
Source: https://ai-powered.com/mcp/mizan-ledger
