Hangar AI Labs

Hangar AI Labs presents

Agent Hangar

Where your agents report for work

One open-source home for every AI agent your company builds, whatever tool built it. Agents are built where people already work, tested before they ship, run in their own containers, and are usable from any platform.

Get startedRead the docsTry the demo

01 · Build where you work

Ask your AI for an agent. It builds, tests and ships it.

The hangar is an MCP server, so Claude, ChatGPT, Codex, Cursor or OpenCode are the builder. They pick a chat agent, a coding harness or a multi-agent team from the objective, reuse what the company already has, write the tests and the user guide, and ship only when the tests pass.

  • MCPbuild from the AI tool you already use
  • Stage → prodevery version tested before it ships
  • Reusenew agents from existing ones, unchanged
The user portal: an agent built by Claude, tested and in productionConnect your AI tool →

02 · Digital employees

Hire an agent for a job, with a human in the loop.

A Digital employee has a job profile, a human manager and an authority level the platform enforces before every action, not the prompt. It works on tasks and issues in the background, pauses on exactly the action that needs a decision, and can get a second opinion from a model of another vendor.

  • Authorityalone, with notice, one or two approvals, never
  • Probationreal tasks in stage before admission
  • Org chartwho reports to whom, people and agents
How Digital employees work →

03 · Governed & measured

Nothing reaches production untested. Every cent is counted.

Immutable versions, stage tests and LLM-as-judge rubrics, four-eyes approval, budgets with a hard stop, DLP, audit and SSO with SCIM. Each agent uses the right reasoning effort for its role, compacts long conversations, and can send simple requests to a cheaper model.

  • Budgetsper company, team, agent or project
  • Auditevery change and decision recorded
  • Costper call, per model, per agent
The execution profile →

04 · Connect everything

Every protocol in, hundreds of tools out.

Each agent answers over OpenAI-compatible, MCP, A2A, ACP and AG-UI, so it works in LibreChat, Slack, VS Code, other agents or inside your own app. Agents reach your systems through connectors (official servers, Docker's catalog and the GitHub MCP Registry) and plugins, with every tool approved one by one.

# any agent, any client
POST /gw/<agent>/v1/chat/completions   OpenAI
POST /gw/<agent>/mcp                    MCP
POST /gw/<agent>/a2a                    A2A
POST /gw/<agent>/agui                   AG-UI, inside your app
Connectors and the MCP registry →

05 · Open & self-hosted

Your models, your servers, Apache-2.0.

The hangar never hosts a model: agents call your LiteLLM, OpenRouter or provider. Run it on a laptop, a cloud VM or Kubernetes (GKE, AKS, EKS) with a kernel sandbox for agents, signed images and an SBOM.

git clone https://github.com/valteresj2/agent-hangar && cd agent-hangar
./scripts/install.sh   # Linux, macOS, WSL
InstallKubernetesStar on GitHub