v1.2.2 — Open Source — MIT License

Your organization runs MCP servers.
Do you control who uses them?

Hangar is the control plane between your AI agents and your MCP servers. One place to decide who can call what, see what's happening, and prove it to your auditors.

Open-source, self-hosted. No cloud account required to get started.

How it works

One proxy between your agents and your MCP servers

AI Agents Claude, GPT, custom
Hangar
Auth Policy Logs
MCP Servers DB, APIs, tools

Quick Start

From install to first parallel call in 2 minutes.

  1. 1

    Install the agent

    Install from PyPI with pip or uv. Works on macOS, Linux, and Windows.

    bash
    pip install mcp-hangar
  2. 2

    Configure MCP servers

    The wizard detects your runtimes, lets you pick MCP servers, and configures Claude Desktop automatically.

    yaml
    # ~/.config/mcp-hangar/config.yaml
    mcp_servers:
      filesystem:
        mode: subprocess
        command: [npx, -y, "@anthropic/mcp-server-filesystem"]
        args: [/Users/you/Documents]
        idle_ttl_s: 300
    
      fetch:
        mode: subprocess
        command: [npx, -y, "@anthropic/mcp-server-fetch"]
        idle_ttl_s: 300
    
      memory:
        mode: subprocess
        command: [npx, -y, "@anthropic/mcp-server-memory"]
        idle_ttl_s: 300
  3. 3

    Start the server

    Launch the agent. MCP servers come online and are ready to serve tool calls.

    bash
    $ mcp-hangar serve
    Starting MCP Hangar v1.2.2...
      ● filesystem  ready  (245ms)
      ● fetch       ready  (189ms)
      ● memory      ready  (156ms)
    → 3 servers ready | 12 tools | parallel execution enabled
  4. 4

    You're ready

    Restart Claude Desktop. Your tools are available with parallel execution, health monitoring, and circuit breakers — all out of the box.

    [ok] 3 servers ready · 12 tools · parallel execution enabled

Capabilities

Everything you need to run MCP in production

Governance

  • Per-caller access policies

    Per-server, per-tool RBAC. Block writes, allow read-only per team.

  • Tool access filtering

    Allow/deny by glob pattern. Block delete_* across the board.

  • Authentication

    API keys with bcrypt, JWT/OIDC, identity propagation to audit logs.

  • Mutator pipeline

    Priority-ordered payload transformations. Audit-only shadow mode for safe rollouts.

Observability

  • Full call logging

    Every tool call logged with caller identity, server, timestamp, and arguments.

  • OpenTelemetry & Prometheus

    Correlation IDs across parallel calls. Traces, metrics, debug batch execution.

  • Cost attribution

    Per-server, per-tool metrics with token, duration, and composite pricing models.

  • Interceptor framework

    Hook-based event model with wildcard subscriptions. Discoverable interceptor sidecar.

Reliability

  • Parallel execution

    Concurrent tool calls across servers. Backpressure and fair scheduling under load.

  • Circuit breaker

    Failing servers isolated automatically. Exponential backoff, graceful degradation.

  • Server groups & load balancing

    Failover, health tracking, round-robin, weighted, and priority-based routing.

  • Lifecycle management

    Lazy loading on first call. Auto shutdown after idle TTL. Health monitoring.

Compliance

  • Audit trail

    Identity-aware logs. SOC2 evidence generation. EU AI Act readiness.

  • SIEM export

    JSONL, CEF, LEEF (QRadar), Syslog (RFC 5424). File, callback, or stderr.

  • Digest pinning

    SHA-256 fingerprint of tool schemas. Detect drift, block mutations, pin versions.

  • Multi-server orchestration

    Local, Docker, remote HTTP — unified interface, heterogeneous backends.

Production hardening

Transport

  • TLS with custom CA support
  • Mutual TLS (mTLS) for agent auth
  • SSRF protection layer

Identity

  • JWT/OIDC with JWKS validation
  • API keys with bcrypt + auto-expiry
  • End-to-end identity propagation

Runtime

  • Capability declaration & verification
  • Exponential rate limiting
  • Command allow-list enforcement

Documentation

Dive deeper into Hangar

Built on Open Source

The mcp-hangar agent is MIT-licensed and always will be. Run it locally, on your servers, or in Kubernetes — no cloud account required. The platform adds managed infrastructure, team collaboration, and enterprise compliance on top.

$ pip install mcp-hangar