01
Interview
A few questions about who this is for, what you do, and how it should talk. No jargon.
A new kind of AI harness
Most AI tools are chat boxes or engineering kits. Saddle interviews you in plain language, then builds a working harness — skills, MCP servers, tools, and connected models — that anyone can use.
01
A few questions about who this is for, what you do, and how it should talk. No jargon.
02
Saddle writes standing orders, skills, and tool catalogs — the same parts engineers wire by hand.
03
Talk to it. Add unlimited skills, MCP servers, and tools as your needs grow. Connect any model.
Why this exists
Powerful. You write code, graphs, or YAML. Built for software people.
Easy. They rarely grow real skills, tools, or servers. They stall when work gets specific.
You answer. It builds the harness. Unlimited skills and MCP tools, usable by a non-engineer.
A model answers. A harness acts. It is the runtime around the model: the loop, the tools it may touch, the skills it can load, the memory it keeps, and the brakes.
In 2026 that layer split in two. Orchestration frameworks (LangGraph, CrewAI) make you draw the workflow. Execution harnesses (Claude Agent SDK, Codex) give a strong model a computer and a loop. Both still assume you can code.
Saddle compiles a harness from an interview. The output is the same anatomy: a system prompt, SKILL.md-style packages, MCP tool catalogs, and a provider mesh (Grok credits or your own keys).
No graphs. No YAML. No “bring your own vector store.” Sign in, answer the questions, talk to the result. Add more skills and servers whenever you want — including by asking the harness to mint them.