for agent developers and tool builders

The hotel is an MCP server, too.

Deploy a guest from your own agent: list, create, redeploy, delete, check health, pull logs. Both surfaces share one endpoint, so a supervising agent can cross-check what a working agent claims against what the host recorded.

One endpoint, two audiences

The panel you click and the endpoint your agent calls are the same installation. Point a client at it and the fleet becomes a tool:

{
  "mcpServers": {
    "agenthotel": {
      "url": "https://panel.yourdomain.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

It is off until you enable it under System → MCP Server. The token is the panel's own — root-equivalent on that host — so treat it accordingly and rotate it from the same page.

What your agent can call

▸

Fleet

list_agents, get_agent, create_agent, delete_agent, redeploy_agent, get_agent_logs.

⌘

Put agents to work

ask_agent hands a hosted agent a task and returns the reply, or dispatches it in the background so it keeps working after the call returns.

✳

Credentials on demand

set_agent_env gives an agent already checked in the credentials its next job needs — stored in its config, so they survive restarts and redeploys.

◉

Observability

health_check, get_host_metrics, get_agent_stats, get_uptime, get_docker_usage, get_events, system_status, run_cleanup.

▤

Templates

list_runtimes, list_templates, get_template, create_template — the same data-driven cards the deploy form is built from.

↻

Between hotels

create_transfer and import_from_hotel move an agent, memory included, server to server.

⚑

Into the container

exec_in_agent runs a command inside an agent's container for inspection. It does not touch the VPS host — that is the panel's own Server Console.

◇

Or skip MCP

Everything the UI does is also a token-authenticated JSON API: agents, stats, uptime, events, templates, providers, domains, certificates, system.

the part that is different

Four panes, four models, one question

Chat opens up to four agents at once — each a live session in its own container, on whatever model that agent runs. Send one message to all of them and read the answers side by side: the quickest way to compare two models, or two harnesses, on a real question.

The panes speak Agent Client Protocol to hermes acp and openclaw acp: streamed answers, thinking, tool calls and plans as they happen, permission requests as buttons. Close the tab and the session detaches instead of ending; the pane comes back to the same process, or restores the conversation in a new one. An idle chat pauses after 30 minutes and picks up where it left off.

Because both surfaces share one endpoint, a supervising agent can cross-check what a working agent claims against what the host recorded.

AgentHotel README · on why the surfaces are shared

What you get to build on

A room per agent — its own container, its own RAM and CPU cap, its own subdomain with automatic HTTPS, its own volumes for sessions, memory and workspace files.
A template that is a recipe — an image, a compose file or a repository, built from its Dockerfile or run with its compose file. Declared secrets are generated at deploy and never shared between two deployments.
A shared table — SkillHub is self-hosted Supabase with pgvector and storage, so guests can find what a colleague already worked out instead of starting from nothing.
A way out — an agent's memory is its harness's own files, and they travel with it. That is what makes the export, the import and the transfer link honest.

Developer questions

How do I test an agent without opening a chat?

There is a test action that sends one message the way the agent's own chat would, with the answer or the agent's own error on screen. A missing provider key shows up in twenty seconds instead of a first failed conversation.

Can I point an agent at my own MCP servers?

Yes. MCP servers are configuration in the panel: point a Hermes agent at one and the block is written into its config on deploy, so a new agent is born able to reach the organisation's tools and an agent recreated from its config comes back whole.

Does a redeploy lose the conversation?

No. Sessions, memory and workspace files live in volumes that outlive the container. Dashboard sessions are signed with a key the panel generates per agent, so redeploying does not log you out.

What if an agent is unhealthy right after deploy?

Give it a minute. Hermes takes about 30 seconds from container start to listening and answers 502 until then; its log says HERMES_DASHBOARD_READY when it is up, and the health sweep corrects itself on the next pass.

Which runtimes can I deploy?

Hermes, OpenClaw, Odysseus, a generic Docker App, a Git App built from a repository, a Git Compose stack, and full Docker Compose deployments.

Every page in this hotel is one markdown file

This is the sites block: a directory of what this Lobby instance serves. Each card below is a single markdown file, written over MCP, with no database and no build step.

Point a client at the endpoint and let it hire the staff.

One panel, one MCP endpoint, one REST API — and a fleet you can inspect from either side.

Read the repository MCP integration