Alpha Harness sits between Claude Code, Codex, Cursor or Gemini CLI and your repo. It stops destructive commands before they run, loads your rules and notes into every session, and keeps your code on your machine.
Illustration. The verdict is the real one, returned by the shipped guard on Apple M2 Max, 2026-10-02.
Works with the tools you already pay for
ClaudeClaude Code
Google GeminiGemini CLI
OpenAICodex
CursorCursor
Google GeminiAntigravity
Terminal toolHermes Agent
Terminal toolAider
Terminal toolDroid
Terminal toolAmp
Terminal toolPi
ClaudeClaude Code
Google GeminiGemini CLI
OpenAICodex
CursorCursor
Google GeminiAntigravity
Terminal toolHermes Agent
Terminal toolAider
Terminal toolDroid
Terminal toolAmp
Terminal toolPi
Why Alpha Harness
Four layers between your agent and your repo.
Each one runs on your machine, each one is tested, and each one tells you exactly what it did.
01
Guards
The command that wrecks a week never runs.
Everything your agent tries to do is checked on your machine first. The destructive ones are refused, with a reason the agent understands, so it takes the safe path instead.
Force-pushes, mass deletes, overwritten secrets and schema pushes stop here.
Normal work is untouched: status, tests and commits to your own branch pass.
Careful mode holds commits, pushes and deploys for a human.
Under 20 ms per check (p95, Apple M2 Max). Works offline.
what your agent tried
git push origin main --force
Blocked
rewrites shared history
rm -rf ./build
Blocked
deletes with no way back
echo "API_KEY=x" > .env
Blocked
overwrites your secrets
npx drizzle-kit push
Blocked
changes a live database
git status
Allowed
npm test
Allowed
git commit -m "wip"
Allowed
Real verdicts from the shipped guard · 2026-10-02
02
Memory
Your agents stop asking what you already decided.
What you settled last month comes back at the moment it matters, and only then. When nothing is relevant, nothing is added: no padding, no bill for noise.
Uses the notes you already keep: Obsidian, Logseq, a plain folder. No notes yet? It starts a vault for you.
Every project sees only its own notes.
Secrets are filtered out before anything is remembered.
Your notes and your code stay on your machine.
what your agent was handed
“why do payouts fail on Base?”2 of your notes recalledPayout retries: what we decidedBase RPC limits
“rename this variable”Nothing relevant. Nothing added.
Illustration. The note titles are examples.
03
Laws
One rulebook. Every tool follows it.
18 laws to start with — check before you claim, no destructive shortcuts, test before "done" — and room for your own. Each tool gets them where it actually reads them, and every copy stays identical.
One install finds the tools you have.
Your own text in those files is kept, never rewritten.
A session that was already open picks the rules up on its next prompt.
Uninstall removes exactly what was added.
what each tool gets
ClaudeClaude Code
rules + guards
OpenAICodex
rules + guards
Google GeminiGemini CLI
rules + guards
Terminal toolHermes Agent
rules + guards
CursorCursor
rules + command guard
Google GeminiAntigravity
rules
Terminal toolAider
rules
The installer's own list. Limits are spelled out on the download page.
04
Doctor
It checks itself, and fixes what is safe to fix.
Protection that quietly stopped working is worse than none. The Doctor tests every layer end to end on a schedule, repairs what is safe to repair, and shows you what it did.
Repairs only touch its own files. Never your notes, never a tool's settings.
Every repair is tested again before it is reported as fixed.
Tells you when something else on the machine is getting in the way.
One click to run it now, one click to copy the report.
doctor · example run
OKGuardsinstalled and answering
OKRulesidentical in every tool
FIXEDMemorysearch index rebuilt
WARNMachineanother harness is also installed
Illustration of one run.
Measured, not promised
Fast enough to forget. Strict enough to trust.
Every check runs on your machine, in milliseconds. No server in the loop, so nothing waits on a network.
20 ms
per guard check, p95
Apple M2 Max
1,355
automated tests
1,036 engine · 319 app
33
guard rules
built in, no setup
18
laws
one rulebook, every tool
10
AI tools covered
one install
Measured 2026-10-02 on the shipped build. Method and history on the changelog.
Trust
Your code, your machine, your audit trail.
Your machine
Guards, memory and the Doctor run locally and keep working offline. Your code and your notes are never uploaded.
Your files
Rules are plain files on your disk. What we add is clearly marked, and your own text is never rewritten.
Your audit trail
Every block is logged with its rule. The Privacy screen prints the exact bytes of anything sent, and every part of it can be switched off.
Your exit
Uninstall removes every file and hook it added, in every tool, and says what it did.
Pricing
Free to start. One price, locked.
The first 100 customers keep this rate for as long as they stay. 30-day money-back guarantee.
Tell us how many people and which tools they use. A number comes back the same day. Or email support@alphaharness.xyz.
30-day money-back guarantee · cancel any time · invoice and bank transfer available · refer someone, earn 30%
Questions
Before you install.
Does it slow my terminal down?
No. Each check is measured at 20 ms at the 95th percentile on an Apple M2 Max. You will not notice it.
Does my code leave my machine?
Never. Rules, notes and guards are files on your disk and work offline. Every plan sends anonymous usage data with no file names, prompts or commands in it, and you can switch any part of it off. Every switch is on the app's Privacy screen, next to the exact bytes.
Which AI coding tools does it support?
Claude Code, Gemini CLI, Codex, Cursor, Antigravity, Hermes Agent, Aider, Droid, Amp, Pi. The download page lists exactly what each one gets, and what it cannot do.
How do I undo it?
In the app: Settings → Uninstall. One click, and every file goes back the way it was. Your notes are never deleted.