mcp-json-yaml-toml

A token-efficient, schema-aware MCP server for safely reading and modifying JSON, YAML, and TOML files

Getting Started • CLI Usage • Available Tools • Development

Stop AI coding tools from breaking your data files. No more grep guesswork, hallucinated fields, or non-schema-compliant data added to files. This MCP server gives AI assistants a strict, round-trip safe interface for working with structured data.

The Problem

AI coding tools often destroy structured data files:

• They grep through huge json, yaml, and toml files (like json logs, or AI transcript files) and guess at keys.
• They hallucinate fields that never existed.
• They use sed and regex that leave files in invalid states.
• They break YAML indentation and TOML syntax.
• They can't validate changes before writing.

The Solution

mcp-json-yaml-toml provides AI assistants with proper tools for structured data:

• Token-efficient: Extract exactly what you need without loading entire files.
• Schema validation: Enforce correctness using SchemaStore.org or custom schemas.
• Safe modifications: Enforced validation on write; preserve comments and formatting.
• Multi-format: JSON, YAML, and TOML through a unified interface.
• Directive-based detection: Support for # yaml-language-server, #:schema, and $schema keys in all formats.
• Constraint-based guided generation: Native LMQL support for proactive validation of partial inputs.
• Local-First: All processing happens locally. No data ever leaves your machine.
• Transparent JIT Assets: The server auto-downloads yq if missing and fetches missing schemas from SchemaStore.org for local caching.

Getting Started

Prerequisites

• Python ≥ 3.11 installed.
• An MCP-compatible client (Claude Code, Cursor, Windsurf, Gemini 2.0, n8n, etc.).

Installation

The server uses uvx for automatic dependency management and zero-config execution.

AI Agents & CLI Tools

uvx mcp-json-yaml-toml

Claude Code (CLI)

claude mcp add --scope user mcp-json-yaml-toml -- uvx mcp-json-yaml-toml

Other MCP Clients

Add this to your client's MCP configuration:

{
  "mcpServers": {
    "json-yaml-toml": {
      "command": "uvx",
      "args": ["mcp-json-yaml-toml"]
    }
  }
}

Schema Discovery & Recognition

The server automatically identifies the correct JSON schema for your files using multiple strategies:

1. Directives: Recognizes # yaml-language-server: $schema=... and #:schema ... directives.
2. In-File Keys: Detects $schema keys in JSON and YAML (also supports quoted "$schema" in TOML).
3. Local IDE Config: Discovers schemas from VS Code/Cursor extension settings and caches.
4. SchemaStore.org: Performs glob-based auto-detection against thousands of known formats.
5. Manual Association: Use the data_schema tool to bind a file to a specific schema URL or name.

LMQL & Guided Generation

This server provides native support for LMQL (Language Model Query Language) to enable Guided Generation. This allows AI agents to validate partial inputs (e.g., path expressions) incrementally before execution.

• Incremental Validation: Check partial inputs (e.g., .data.us) and get the remaining pattern needed.
• Improved Reliability: Eliminate syntax errors by guiding the LLM toward valid tool inputs.
• Rich Feedback: Get suggestions and detailed error messages for common mistakes.

Available Tools

Development

Setup

git clone https://github.com/bitflight-devops/mcp-json-yaml-toml.git
cd mcp-json-yaml-toml
uv sync

Testing

ash

Run all tests (coverage included)

uv run pytest

### Code Quality

The project uses `prek` (a Rust-based pre-commit tool) for unified linting and formatting. AI Agents MUST use the scoped verification command:

```bash
# Recommended: Verify only touched files
uv run prek run --files <file edited>

Project Structure

mcp-json-yaml-toml/
├── packages/mcp_json_yaml_toml/  # Core logic
│   ├── server.py                 # MCP implementation
│   ├── yq_wrapper.py             # Binary management
│   ├── schemas.py                # Schema validation
├── .github/                      # CI/CD and assets
├── docs/                         # Documentation
└── pyproject.toml                # Project config

# Run all tests (coverage included)
uv run pytest

Code Quality

The project uses prek (a Rust-based pre-commit tool) for unified linting and formatting. AI Agents MUST use the scoped verification command:

# Recommended: Verify only touched files
uv run prek run --files <file edited>

Project Structure

graph TD
    Repo[mcp-json-yaml-toml]
    Repo --> Packages[packages/mcp_json_yaml_toml]
    Repo --> Github[.github]
    Repo --> Docs[docs]
    Repo --> Config[pyproject.toml]

    subgraph "Core Logic"
        Packages --> Server[server.py<br/>MCP Server & Tools]
        Packages --> Schemas[schemas.py<br/>Schema Validation]
        Packages --> Constraints[lmql_constraints.py<br/>LMQL Constraints]
        Packages --> YQ[yq_wrapper.py<br/>Binary Manager]
        Packages --> YAML[yaml_optimizer.py<br/>YAML Anchors]
        Packages --> TOML[toml_utils.py<br/>TOML Utils]
        Packages --> Conf[config.py<br/>Config Manager]
    end

    style Packages fill:#f9f,stroke:#333,stroke-width:2px
    style Repo fill:#eee,stroke:#333,stroke-width:4px

Token Efficiency Experiment

Two identical Claude Code sub-agents were given the same task: read ~/.claude.json and report every MCP server listed, including command, args, and env vars.

Setup

• Agent A — standard prompt, used the built-in Read tool
• Agent B — same prompt with one line appended: You must use the mcp__json-yaml-toml for all file interactions.

Both agents used the sonnet model.

Prompts

Agent A prompt:

Read the file ~/.claude.json and report back:
1. Every MCP server listed in the mcpServers section
2. For each server: the command, args, and any env vars configured

Just report the raw findings. Do not summarize or interpret.

Agent B prompt:

Read the file ~/.claude.json and report back:
1. Every MCP server listed in the mcpServers section
2. For each server: the command, args, and any env vars configured

You must use the mcp__json-yaml-toml for all file interactions.

Just report the raw findings. Do not summarize or interpret.

Results

Both agents returned identical findings (8 MCP servers with correct configs).

Agent B used 22.6% fewer tokens and completed in 43% of the time with half the tool calls.

Why

The Read tool loads the entire file into context. ~/.claude.json is a large file — the agent had to consume all of it to find the mcpServers section. The MCP server's data_query tool extracted just the mcpServers section directly, keeping the context window small.

Built with FastMCP, yq, and LMQL