Agent harnesses drift when every repository invents its own setup.
Harness Kit exists for teams that need AI coding guardrails to be repeatable, reviewable, and owned by the project rather than hidden inside a personal prompt folder.
Setup drift
Rules, skills, hooks, MCP servers, and agent instructions diverge across repositories when they are copied by hand.
Context budget pressure
Every extra rule consumes model context. Bundles need to justify their cost instead of dumping everything into every project.
Unreviewable installation
One-off setup changes are hard to audit. Teams need small diffs, explicit files, and status checks.
Install one capability, preview the artifacts, verify the state.
The CLI keeps the workflow intentionally mechanical: initialize, select a bundle, apply owned files, then audit the harness state from the repository itself.
Previewed artifacts
Initialize
Detect project shape, ask the minimum setup questions, and write the baseline harness files.
Select bundle
Choose workflow, stack, tech-stack, memory, search, browser, or MCP capability from the registry.
Preview diff
Show which files will be written so the harness can enter the project through a normal code review.
Audit status
Report installed bundles, missing files, env vars, and drift before the next agent run depends on them.
A typed registry feeds a copy-owned installer.
The system is deliberately layered: shared types define the contract, the registry describes bundles as data, the engine writes artifacts, and commands/TUI orchestrate user decisions.
- 01
Core contract
Shared TypeScript types define bundle manifests, artifact variants, env vars, and harness config so every layer speaks the same schema.
@harness-kit/coretypesconstants - 02
Bundle registry
Bundles are TypeScript manifests, not loose JSON. The catalog can be queried, validated, and rendered into web docs.
BundleManifestworkflowtechstack - 03
Installer engine
Artifact installers and template renderers write files into the target repo while preserving an auditable result.
installertemplatesscaffolder - 04
Command and TUI layer
Commander commands and an Ink/XState wizard coordinate init, add, list, status, and preview/apply flows.
commanderInkxstate
Project-owned guardrails, installed in small units.
Harness Kit focuses on concrete repository artifacts instead of a hidden runtime service.
Small, reviewable diffs
Add one capability at a time so setup changes stay understandable and easy to audit.
Consistent team baseline
Standardize agent rules, skills, and harness state across repositories.
Practical onboarding docs
Follow quickstart, command reference, and troubleshooting without reading source code first.
The toolkit optimizes for ownership, not invisible automation.
Each design choice keeps the harness inside the repository, even when that means the user must review files and manage context deliberately.
- 01
Copy-own distribution
Optimizes
Optimized for editable, versioned artifacts with no runtime dependency on the CLI after installation.
Accepted cost
Updates are not automatic. Future changes need explicit update flows and reviewable diffs.
- 02
Just enough context
Optimizes
Optimized for keeping agent instructions below the point where context cost harms output quality.
Accepted cost
The default preset must stay opinionated and cannot include every useful rule by default.
- 03
Registry as typed data
Optimizes
Optimized for validation, docs generation, and agent-readable bundle metadata.
Accepted cost
Adding a bundle requires disciplined manifest structure instead of dropping arbitrary files into a folder.
A published beta CLI with a growing bundle catalog.
Harness Kit already has the shape of a real developer tool: package boundaries, command tests, a TUI wizard, registry validation, web docs, examples, and release planning.
CLI and wizard are usable
The beta CLI supports init, add, list, status, budget, version, activate, and removal flows around project-owned harness files.
Bundle catalog exists
Workflow, stack, and tech-stack bundles cover skills, rules, agents, hooks, MCP integrations, and framework guidance.
Next evidence is usage telemetry
The planned usage command should show which skills and agents are useful, stale, or damaging context efficiency.