Vibeleaderboard
Index / app
Visit github.com
Category
Developer Tools
Rank

Previous survey · No. 844 ·

Type
APP
Latest release
v5.15.0
Date

About

A dedicated development harness for Claude Code aimed at high-quality, repeatable output.

What it does

Claude Code Harness turns agent coding into a gated delivery loop. It drafts an explicit contract, limits implementation to approved work, verifies results, runs a separate review, checks drift, and packages release evidence. A Go engine also evaluates tool activity before execution and records blocked actions without storing command text.

Why it's ranked here

The design addresses real failure modes: disappearing plans, optional tests, late review, and reconstructed release evidence. Its strongest distinction is enforceable procedure, including non-overridable safety categories and scoped approvals. The verdict is tempered by uneven protection across supported hosts and documentation that sometimes reflects different architectural eras.

What's good

The workflow produces durable handoffs between planning, implementation, review, and release. Unknown facts remain explicitly unknown. Major review findings block completion. Risky operations can be approved at planning time with expiry, scope, and usage limits. Non-engineers also receive compact plan, progress, and acceptance views instead of needing to inspect code.

Tradeoffs

Protection varies by host. Claude Code can stop many tool types before execution, while Codex only has strong pre-execution coverage for shell activity. Cursor lacks a traditional filesystem jail and may fail open in some cases. Grok is outside the shared safety floor. Automatic approval remains experimental and does not yet skip prompts.

How to use it well

Use it for repository work where auditability, controlled scope, independent review, and release evidence justify extra ceremony. Start with a small task, correct the generated contract, then approve execution. Prefer the Claude Code route for safety-sensitive changes. It does not improve model intelligence, replace human approval, or provide equal containment across every host.

Technical notes+

README.md defines five core verbs plus setup, a two-layer safety model, JSONL stop logging, and marketplace or script-based installation. docs/MIGRATION-v4.md says hooks now invoke the bundled bin/harness Go binary directly, replacing the former TypeScript engine and Node.js runtime; docs/claims-audit.md still labels the TypeScript guardrail engine proven, so that audit appears architecturally stale. docs/hardening-parity.md documents the shared pre-tool kernel and its host asymmetries. docs/i18n.md defines English and Japanese output, with message instructions taking precedence over project configuration and environment settings. docs/ARCHITECTURE.md describes skill, workflow, and profile layers, while docs/team-composition.md specifies Lead-controlled workers, an optional advisor, a serial reviewer verdict, and worktree isolation.

Observed

License
MIT
Core runtime
Go-native guardrail engine with no Node.js requirement
Install surface
Claude Code plugin marketplace, plus setup scripts for Codex CLI, Cursor, and Grok
Interfaces
Slash-command skills and a bundled command-line binary
Supported hosts
Claude Code, Codex CLI, Cursor, and Grok are listed as supported, with different enforcement guarantees
Platforms
Prebuilt binaries for macOS arm64 and amd64, Linux amd64, and Windows amd64
Languages
English is the default user-facing language; Japanese is also supported

Read from README.md, docs/i18n.md, docs/ARCHITECTURE.md, docs/MIGRATION-v4.md, docs/claims-audit.md, docs/MEMORY_POLICY.md, docs/content-layout.md, docs/CLAUDE-commands.md, docs/judgment-ledger.md, docs/opus-4-7-impact.md, docs/advisor-strategy.md, docs/benchmark-rubric.md, docs/hardening-parity.md, docs/team-composition.md, docs/agent-view-policy.md.

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.