MCP Hangar

Kubernetes-native registry for managing multiple MCP servers with lazy loading, health monitoring, and RBAC

MCP Hangar

PyPI Python 3.11+ License: MIT

Production-grade infrastructure for Model Context Protocol.

MCP Hangar is a control plane for MCP servers. It manages MCP server lifecycle, parallel tool execution, security governance, and observability -- so you don't have to.

Quick Start

30 seconds to working MCP servers:

curl -sSL https://mcp-hangar.io/install.sh | bash && mcp-hangar init -y && mcp-hangar serve

That's it. Filesystem, fetch, and memory MCP servers are now available to Claude.

What just happened?
  1. Install - Downloaded and installed mcp-hangar via pip/uv
  2. Init - Created ~/.config/mcp-hangar/config.yaml with starter MCP servers
  3. Serve - Started the MCP server (stdio mode for Claude Desktop)

The init -y flag uses sensible defaults:

  • Detects available runtimes (uvx preferred, npx fallback)
  • Configures starter bundle: filesystem, fetch, memory
  • Runs a smoke test to verify MCP servers start correctly
  • Updates Claude Desktop config automatically

Manual Setup

# 1. Install
pip install mcp-hangar
# or: uv pip install mcp-hangar

# 2. Initialize with wizard
mcp-hangar init

# 3. Start server
mcp-hangar serve

HTTP Mode

# Start with HTTP transport and REST API
mcp-hangar serve --http --port 8000

# REST API:  http://localhost:8000/api/

What It Does

Parallel execution. Your AI agent calls 5 tools sequentially -- each takes 200ms, that's 1 second of waiting. hangar_call runs them in parallel. 200ms total.

hangar_call(calls=[
    {"mcp_server": "github", "tool": "search_repos", "arguments": {"query": "mcp"}},
    {"mcp_server": "slack", "tool": "post_message", "arguments": {"channel": "#dev"}},
    {"mcp_server": "internal-api", "tool": "get_status", "arguments": {}}
])

Single MCP tool call. Parallel execution. All results returned together.

Lifecycle management. Lazy loading, health checks, automatic restart, graceful shutdown. MCP servers start on first use, stay warm while active, shut down after idle TTL.

Single-flight cold starts. When 10 parallel calls hit a cold MCP server, it initializes once -- not 10 times.

Circuit breaker. One failing MCP server doesn't kill your batch. Automatic isolation and recovery.

Configuration

mcp_servers:
  github:
    mode: subprocess
    command: [uvx, mcp-server-github]
    env:
      GITHUB_TOKEN: ${GITHUB_TOKEN}

  slack:
    mode: subprocess
    command: [uvx, mcp-server-slack]

  internal-api:
    mode: remote
    endpoint: "http://localhost:8080"

  custom-server:
    mode: docker
    image: my-registry/mcp-server:latest
    container:
      command: ["python", "-m", "custom_entrypoint"]

Claude Desktop Integration

mcp-hangar init auto-configures Claude Desktop. For manual setup, add to your Claude Desktop config:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Linux: ~/.config/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "hangar": {
      "command": "mcp-hangar",
      "args": ["serve", "--config", "~/.config/mcp-hangar/config.yaml"]
    }
  }
}

Restart Claude Desktop. Done.

Python API

For programmatic use (scripts, pipelines, custom integrations):

from mcp_hangar import Hangar, HangarConfig

# Async
async with Hangar.from_config("config.yaml") as hangar:
    result = await hangar.invoke("math", "add", {"a": 1, "b": 2})

# Sync wrapper
from mcp_hangar import SyncHangar

with SyncHangar.from_config("config.yaml") as hangar:
    result = hangar.invoke("math", "add", {"a": 1, "b": 2})

# Programmatic config
config = (
    HangarConfig()
    .add_mcp_server("math", command=["python", "-m", "math_server"])
    .add_mcp_server("fetch", mode="docker", image="mcp/fetch:latest")
    .build()
)
hangar = Hangar(config)

Security & Governance (1.0)

  • Capability declaration. Declare what each MCP server can access (network, filesystem, environment). Violations are detected and reported.
  • Behavioral profiling. Baseline MCP server behavior, detect deviations (new destinations, protocol drift, frequency anomalies). Learning and enforcing modes.
  • Tool schema drift detection. Track tool schema changes across MCP server updates.
  • Network connection monitoring. /proc/net/tcp parsing, Docker and Kubernetes monitors with audit events.
  • RBAC. Role-based access control with tool-level policies. API key and JWT/OIDC authentication.
  • Approval gate. Human-in-the-loop approval for sensitive tool calls.

Observability

  • OpenTelemetry. Distributed tracing with W3C trace context propagation across MCP servers.
  • Prometheus metrics. MCP server state, tool calls, health checks, circuit breaker, concurrency, batch execution.
  • Grafana dashboards. Pre-built overview and per-MCP server deep dive dashboards.
  • Structured logging. Correlation IDs across parallel calls. JSON log format for production.
  • Audit trail. Event-sourced audit log with OTLP export for security-relevant events.

Advanced Configuration

mcp_servers:
  fast-mcp-server:
    mode: subprocess
    command: ["python", "fast.py"]
    idle_ttl_s: 300              # Shutdown after 5min idle
    health_check_interval_s: 60  # Check health every minute
    max_consecutive_failures: 3  # Circuit breaker threshold
    max_concurrency: 5           # Per-MCP server concurrency limit
    tools:
      deny_list: [delete_*]      # Tool access filtering

execution:
  max_concurrency: 50            # Global concurrency limit
  default_mcp_server_concurrency: 10

truncation:
  enabled: true
  max_batch_size_bytes: 950000   # Under Claude's 1MB limit

config_reload:
  enabled: true                  # Live config reload via file watch

Scales With You

  • Home lab: 2 MCP servers, zero config complexity
  • Team setup: Shared MCP servers, Docker containers, hot-reload
  • Enterprise: 50+ MCP servers, behavioral profiling, RBAC, approval gates, Kubernetes operator

Same API. Same reliability. Different scale.

Documentation

License

Core (src/) is MIT licensed. Enterprise features (enterprise/) are BSL 1.1 licensed.

See LICENSE for MIT terms and enterprise/LICENSE.BSL for BSL terms.


Docs | PyPI | GitHub

เซิร์ฟเวอร์ที่เกี่ยวข้อง

NotebookLM Web Importer

นำเข้าหน้าเว็บและวิดีโอ YouTube ไปยัง NotebookLM ด้วยคลิกเดียว ผู้ใช้กว่า 200,000 คนไว้วางใจ

ติดตั้งส่วนขยาย Chrome