Skip to content

ai_stp

ai_stp helps a developer and their coding agent select, verify and safely install a complete setup for an AI harness.

A setup covers instructions, skills, MCP, hooks, commands, agents, plugins and settings. ai_stp records provenance, compatibility, exact versions and trust decisions so that an agent never has to guess at a configuration.

The website owns the account and the public catalog. It displays results. It does not select a composition, assemble a setup, or write native harness state. That work belongs to the CLI, the agent, and the harness's own public provider.

Where to start

  • Quickstart for people: install the CLI, check the environment and read the catalog.
  • Quickstart for agents: start every session from task intents --json.
  • CLI: the working surface — JSON envelopes, plans, and confirmation.
  • Web: account, catalog cards, publications, and reports.
  • Harnesses: the seven open-beta harnesses and undefined.
  • Concepts: harness, setup, provider, assembler, device, project, and the three modes.
  • Components: the closed kinds and how they differ.
  • Catalog: how to read a public result and how the CLI searches it.
  • Trust and safety: why a verified author is not the same thing as safe content.
  • Troubleshooting: recovering when an install or a check does not pass.

What the MVP supports

Open beta ships all seven harnesses at the beta tier: Claude Code, Codex, Grok Build, Pi, OpenCode, Cursor, and Antigravity. Missing evidence is not_verified; it does not drop a harness from the product.

An unknown harness falls back to limited undefined. Reading, import and local checks are allowed. Automatic installation is not considered safe.

The main path looks like this:

CLI → task intents → task start
→ follow continuations (execute argv only when actor is cli)
→ report payload verification
How to read this documentation

If ai_stp is new to you, start with the human or agent quickstart and the harnesses page. If you are already assembling a setup, go straight to components: each page explains what one kind is for, where its boundary is, and what it risks. Everyday flags come from continuation argv after ai-stp task intents --json; help --agent remains the full registry of this install. This site names commands so a person can find the right page.