Where agentsdeploy agents.
Every agent has a remote of its own. Push to it and the rest runs without you.
Deployment stops at a git push.
$ git push safehouse main
live at https://safehouse.run/a/x7f2
Five formats, none of them privileged. Whichever yours is written in, it runs here as it is.
We support the agents you already built.
Filesystem-first. Instructions, tools, skills, channels and schedules are all just files in a directory.
The agent harness as a framework. Headless and programmable, so it runs with nobody present.
Flue rebuilt on hooks. useModel, useTool, useSkill — agents that change what they are at runtime.
Stateful agents as a TypeScript class, each one its own Durable Object with memory and scheduling.
Primitives for long-running agents, and an opinionated base class that wires them together.
Point your agent at our MCP. It has the tools to do the rest.
Your agent does the setup.
connect safehouse-mcp
6 tools, scoped to your account
create_agent { name: "support-triage" }
agent_x7f2
git remote safehouse.run/git/agent_x7f2.git
push token issued once, shown once
writes the agent directory
git push safehouse main
live at https://safehouse.run/a/x7f2
A directory on your machine becomes an agent on the edge.
Get access.
Invite-only while we build it out. Leave your email and we’ll tell you when there is room.