Skip to content

horsie

An open-source managed agent harness. Durable sessions, a sandbox per session, and unattended runs — on infrastructure you own, with models you already pay for.

An agent is a model plus a harness — the loop that assembles context, calls the model, decides whether a tool call is allowed, runs it, and writes down what happened. horsie is the harness. You bring the model.

Run one

Start the server, add a model, give it a sandbox, and hold a session. Quickstart

Sessions that survive

An append-only record of every model call, tool call and result. Close the tab, restart the server, walk back in. Sessions & durability

A sandbox per session

Your own machine over a connection it dials out, or a container the server builds and throws away. Runtimes & vendors

Work without you

Scheduled and API-triggered runs that report back a session you can read and continue — not a log file. Routines

Your own outer harness

Skills, hooks, MCP servers, agent presets and environments, selected per session. The harness

Read the loop

The interesting part of an agent is not the model. All of it is source you can read and change. What horsie is