Agent Harness SDK

Agent Harness SDK wraps Claude Code in a deterministic shell — three TypeScript primitives, enforced by Claude Code's hooks:

• Guards — pre-action policies. Block bad calls before they run.
• Checks — post-action validators. Catch problems and feed them back.
• Tools — deterministic MCP operations. Replace error-prone prose steps.

See How it works for the agentic-loop walkthrough.

Getting Started

Install

npm install -D agent-harness-sdk
npx harness init

Restart Claude Code. You will now have a harness/ directory, hooks registered under .claude/, and the harness server merged into .mcp.json.

Unlock the harness

The harness is locked by default — it protects its own files (harness/, the hook wiring) from the agent, so the agent can't edit or scaffold harness primitives until you unlock it:

npx harness security 0

Scaffold a primitive

From inside Claude Code (the harness must be unlocked first), describe the guard, check, or tool you want in natural language:

/harness add guard to block imports from internal/ outside its module
/harness add check that changed components have a test
/harness add tool to run typecheck and return the errors

The agent names it, scaffolds a typed stub in the right directory, registers it in harness/harness.config.ts, and works with you to write the guard, check, or tool — asking for specifics if your description needs them.

When you're done working on the harness, re-lock it with npx harness security 1. See Security to learn more.