agent-harness
The missing package manager for MCP servers and AI skills
I'm a solo developer based in Islamabad. In May I published the first release candidate of
agent-harness — a CLI for managing MCP servers, rules files, and AI agent skills across
Cursor, Claude Code, Copilot, Zed, and a handful of other editors. Yesterday I shipped
v2.1.1. Here's the honest version of what happened in between.
The problem I kept hitting
Every AI editor has its own config format. Cursor wants .cursor/rules. Claude Code wants
CLAUDE.md and .claude/. Copilot has its own instructions file. If you use more than one
editor — or work across machines — you either copy-paste and pray, or you maintain the same
context in four different formats by hand.
MCP servers made it worse. Each one has to be registered separately per editor, with no
pinning, no reproducibility, no way to know if the version you're running is the same as
last week.
I kept rebuilding my setup from scratch. So I wrote a tool to stop doing that.
What I built
agent-harness is a supply-chain CLI: discover assets from trusted sources, rank them by
authority signals, mirror them locally, stage them for review, activate them, wire them into
whichever editor you're in. The quarantine system — where anything that looks like a
prompt-injection risk gets held for explicit review before it touches a workspace — is the
part I'm most proud of. I haven't seen it in any other tool in this space.
It supports 7 hosts. It has one production dependency. It ships with 100% test coverage
across statements, branches, functions, and lines. Whether or not anyone uses it, I wanted
it to be the kind of software I'd trust.
Distribution: what I tried
Everything you'd expect, and most of it didn't work.
Two Dev.to articles. Zero comments. Medium and Hashnode got similar silence. A "Ask HN"
post was flagged within hours — the framing was too abstract, and HN correctly identified
it as a product promotion dressed up as a question. My Reddit account got suspended during
the outreach period via API-detection; appeals are still ongoing months later. I had a cold
X account with three tweets and no retweet surface.
The CoderLegion posts got views from people who weren't developers. The blog posts got reads
from people who weren't feeling the pain.
What actually worked
Awesome-list PRs. Several have now merged — including into the OpenCode ecosystem and
harness-engineering directories — and the traffic they generate is small but warm. People
who land on agent-harness via an awesome-list are actively looking for what it does.
The YouTube demo helped too, mostly by giving me something credible to link in the GitHub
README above the fold.
Product Hunt is live. The listing is up. The actual launch post goes out on Tuesday.
Where it stands now
v2.1.1 on npm. 14 releases since the first RC on May 5th. ARD-compatible — the Agentic
Resource Discovery spec (Google, Microsoft, Hugging Face, and others) was published earlier
this year and explicitly defers the lifecycle problem to "backend implementation." That
sentence is a clean description of what agent-harness does.
No viral moment. No spike. A small, slow accumulation of the right kind of attention from
people who actually feel the problem.
If you use multiple AI editors and manage MCP servers by hand: npm i -g [@ar27111994](/ar27111994)/agent-harness
If you've been through this kind of distribution slog and have something to share — I'm listening.
5 Comments
The awesome-list traffic sounds small but unusually qualified. Are those visitors actually installing and continuing to use agent-harness, or are you seeing warm clicks without much downstream adoption yet?
I'd say it's mostly warm clicks without much downstream adoption yet.
That distinction between warm clicks and actual adoption is useful. If you’re open to it, what’s the best email to reach you on?
admin@ar27111994.dev
Thanks! I’ve just sent it over.
Looking forward to hearing your thoughts whenever you have a chance.