MCP 服务器

Quilt MCP Server

Enables searching, analyzing, and managing data packages in a Quilt data catalog through natural language, with 84+ tools including package CRUD operations.

README

Quilt MCP Server

MCP server for Quilt data catalog - search, analyze, and manage data packages with 84+ tools.

Quick Start

1. Terminal (Direct)

# Run directly with uvx (requires uv: https://docs.astral.sh/uv/)
# Default deployment is "local" (platform + stdio)
uvx quilt-mcp

# Other deployment modes
uvx quilt-mcp --deployment remote  # platform + http
uvx quilt-mcp --deployment legacy  # quilt3 + stdio

# Or install globally
uv tool install quilt-mcp
quilt-mcp

2. Claude Desktop (One-Click)

Download .mcpb from releases
Double-click to install or drag to Claude Desktop
Configure catalog in Settings → Extensions → Quilt MCP

3. Claude Code CLI

# Add to Claude Code CLI with environment variables
npx @anthropic-ai/claude-code mcp add quilt-mcp uvx quilt-mcp \
  -e QUILT_CATALOG_URL=https://your-catalog.quiltdata.com \
  -e AWS_PROFILE=your-profile

4. Custom MCP Clients

Add to your mcp.json:

{
  "mcpServers": {
    "quilt": {
      "command": "uvx",
      "args": ["quilt-mcp"],
      "env": {
        "QUILT_CATALOG_URL": "https://quilt-stack.yourcompany.com"
      }
    }
  }
}

Configuration

Configure or refresh quilt3 credentials with:

# Configure catalog and authenticate (interactive)
quilt3 config

# Or set directly
quilt3 config https://your-stack.your-company.com

# Login (opens browser for SSO, or prompts for credentials)
quilt3 login

By default, quilt-mcp uses the local deployment mode (--deployment local), which uses the platform backend and requires:

QUILT_CATALOG_URL
authentication (quilt3 login session)

Use deployment presets:

uvx quilt-mcp --deployment remote  # platform + http
uvx quilt-mcp --deployment local   # platform + stdio (default)
uvx quilt-mcp --deployment legacy  # quilt3 + stdio

QUILT_DEPLOYMENT env var can set the same modes:

QUILT_DEPLOYMENT=remote uvx quilt-mcp
QUILT_DEPLOYMENT=local uvx quilt-mcp
QUILT_DEPLOYMENT=legacy uvx quilt-mcp

Remote Docker + ngrok testing hack (for Claude.ai without OAuth):

make run-docker-remote
# Starts Docker container on localhost:8000 with auto-injected real JWT fallback
# Launches MCP Inspector at http://127.0.0.1:6274 for testing
# JWT discovery priority: PLATFORM_TEST_JWT_TOKEN -> quilt3 login session
#
# In another terminal, expose via ngrok:
ngrok http 8000 --domain=$NGROK_DOMAIN

Set NGROK_DOMAIN in .env, and configure Claude MCP URL as https://<your-ngrok-domain>/mcp. Use this only for local development/testing.

Backward compatibility:

--backend still works as an explicit backend override.
QUILT_MULTIUSER_MODE is still supported as a legacy selector.

See docs/AUTHENTICATION.md for full configuration details and examples.

Docker Deployment

Docker images are published to AWS ECR (account 730278974607) with multiple tags for flexibility:

Available tags per release:

Semver tag: 0.17.3 - Human-readable version
Git SHA tag: abc12345 - Git commit hash (8 chars) for traceability
Latest tag: latest - Points to most recent production release

Example usage:

# Pull by version (recommended for production)
docker pull 730278974607.dkr.ecr.us-east-1.amazonaws.com/quiltdata/mcp:0.17.3

# Pull by git commit SHA (for debugging/rollback)
docker pull 730278974607.dkr.ecr.us-east-1.amazonaws.com/quiltdata/mcp:abc12345

# Pull latest (convenient but not recommended for production)
docker pull 730278974607.dkr.ecr.us-east-1.amazonaws.com/quiltdata/mcp:latest

Trace deployed image back to source:

# Get SHA tag from running container (replace CONTAINER_ID with actual container ID)
docker inspect CONTAINER_ID | grep "quiltdata/mcp:"

# View commit details (replace abc12345 with actual SHA from above)
git log --oneline abc12345

See GitHub releases for available versions and their git commit SHAs.

Environment Variables

Override defaults via environment or MCP config:

QUILT_CATALOG_URL - Your Quilt catalog URL (e.g., https://your-catalog.quiltdata.com)
QUILT_DEPLOYMENT - Deployment mode (remote, local, legacy)
QUILT_MULTIUSER_MODE - Legacy backend selector (true -> platform, false -> quilt3)
AWS_PROFILE - AWS credentials profile for S3 access (if not default)
QUILT_SERVICE_TIMEOUT - HTTP timeout for service calls in seconds (default: 60)

Architecture

Multiuser Mode (Production)

Stateless: No server-side workflows or templates
JWT auth: Catalog-issued JWTs only (claims: id, uuid, exp)
Read/write operations go through the catalog API
Horizontally scalable: any number of containers
Single tenant per deployment (no tenant tracking)

Local Dev Mode

Stateful: File-based storage in ~/.quilt/
IAM auth: Uses AWS credentials or quilt3 session
Full feature set, including workflows
Single-user development and testing

Core Package Tools

Tool	Operation	Backend path
`package_create`	Create package revision from S3 objects	`QuiltOps.create_package_revision()`
`package_update`	Update existing package revision	`QuiltOps.update_package_revision()`
`package_delete`	Delete package revisions	`QuiltOps.delete_package()`

Development

# Clone and setup
git clone https://github.com/quiltdata/quilt-mcp-server.git
cd quilt-mcp-server

# Install and run
uv sync
make run

# Test
make test
make test-func
make test-e2e

Testing Infrastructure

The Quilt MCP Server includes a comprehensive testing framework (quilt_mcp.testing) for automated test generation and execution:

Automatic Test Generation: Discovers tools, infers arguments, generates YAML configurations
Intelligent Classification: Categorizes tools by effect (create/update/remove) and requirements
Tool Loop Execution: Multi-step workflows for testing write operations (create → modify → verify → cleanup)
Comprehensive Validation: Result validation, coverage analysis, failure pattern detection

Quick Start:

# Generate test configuration
make test-mcp-setup

# Run all MCP tests
make test-mcp

# Run specific test suites
uv run scripts/mcp-test.py --tools          # Test tools only
uv run scripts/mcp-test.py --resources      # Test resources only
uv run scripts/mcp-test.py --loops          # Test tool loops only
uv run scripts/mcp-test.py --idempotent-only # Test read-only operations

# Run with selectors
uv run scripts/mcp-test.py --tools-select "bucket_list,package_list"

Module Structure:

src/quilt_mcp/testing/ - Testing framework library (4,644 lines)
scripts/mcp-test.py - Test execution script (1,599 lines)
scripts/mcp-test-setup.py - Test generation script (302 lines)

See Testing Framework Documentation for detailed API documentation and usage patterns.

Documentation

Troubleshooting

SyntaxWarning from jsonlines

You may see this warning during installation:

SyntaxWarning: invalid escape sequence '\*'

This is harmless. It's from the jsonlines dependency (via quilt3) and doesn't affect functionality. The warning appears on Python 3.12+ due to deprecated escape sequences in the library's docstrings.