#196196th of 236 in Coding agents and AI editors

metaharness

Generator that scaffolds a repo-specific agent harness as an npm package

Stars
753
License
MIT
Last commit
Nov 2026
Last release
Nov 2026
Language
TypeScript
Stars, last 22 weeks+737
13 Jun 2026: 16 stars20 Jun 2026: 264 stars27 Jun 2026: 337 stars4 Jul 2026: 363 stars11 Jul 2026: 464 stars18 Jul 2026: 487 stars25 Jul 2026: 516 stars1 Aug 2026: 536 stars8 Aug 2026: 561 stars15 Aug 2026: 590 stars22 Aug 2026: 602 stars29 Aug 2026: 623 stars5 Sept 2026: 631 stars12 Sept 2026: 649 stars19 Sept 2026: 665 stars26 Sept 2026: 672 stars3 Oct 2026: 685 stars10 Oct 2026: 694 stars17 Oct 2026: 707 stars24 Oct 2026: 717 stars31 Oct 2026: 738 stars7 Nov 2026: 745 stars14 Nov 2026: 753 stars

16 on 13 Jun 2026753 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

MetaHarness generates an npm-publishable agent harness for a repo from a GitHub URL or a template, using the `npx metaharness` CLI or a browser Studio. The output includes agents, skills, slash commands, an MCP server with a default-deny policy, scoped memory and signed provenance. It targets ten hosts, including Claude Code, Codex, Copilot, OpenCode and GitHub Actions, and has 19 vertical templates.

Who it is for: Teams making repo-specific agent harnesses to publish as internal npm packages

Strengths

  • One scaffold targets ten hosts, including Claude Code, Codex, Copilot, OpenCode and GitHub Actions
  • MCP output is default-deny: no network, no shell, no file-write, audit on
  • `score` and `analyze-repo` read a repo without executing it
  • `mcp-scan` statically flags shell/network grants and wildcard permissions in agent tools

Weaknesses

  • It generates harnesses rather than running agents, so you still need a host and a model
  • Several packages (ARC-AGI-3, AVO) are private or experimental, with performance claims unproven
  • README is long and dense with ADR references, so the core workflow is hard to find
  • Darwin Mode evolution is validated mainly on SWE-bench Lite; broader gains are not shown

What it needs

  • no GPU
  • Needs Node.js, npm
  • Models: Model-agnostic, GLM/Qwen LoRA tuning via weight-eft (GPU-gated)

What changed

From the release notes. No flag means the note does not say so, not that the upgrade is safe.

Also in Coding agents and AI editors

See all 236
Also in Coding agents and AI editors
RankProjectScore
1claw-codeRust CLI coding agent harness using Anthropic and OpenAI-compatible APIs194.8k starsMIT96 out of 100
2t3codeDesktop, web and mobile front end for controlling local coding agent CLIs Live demo ↗ (opens in a new tab)64.4k stars+36,149 (+128%) in 30 daysMIT96 out of 100
3codewhaleTerminal coding agent in Rust that works with 40+ model providers41.2k starsMIT94 out of 100
#4gooseRust AI agent with desktop app, CLI and API, extended through MCP56.7k stars+1,476 (+2.7%) in 30 daysApache-2.093 out of 100
#5ArchonYAML workflow engine that runs AI coding agents through fixed dev steps23.9k starsMIT93 out of 100