Ghost Architect™ Local · version 8.0.58 · early access beta

Version 8, early access

Version 8 is a beta, and we would value your feedback on it. Fill in the form and your download link arrives by email in about a minute. The user's manual is inside the download. The evaluation trial starts inside the product; no card, no account.

Only a few days remain to join the version 8 beta program.

What version 8 adds

Load Qualifying

Ghost Architect™ Local reads your code and shows where it is at risk of failing under load: N+1 queries, unpaginated collections, a page made uncacheable by one block, outside calls with no timeout and more. Each risk comes from a catalogue of 134 known load failures for Magento 2 and Adobe Commerce, PHP, Laravel, Node.js, Python, Java and .NET, with the vendor's own sources. It never runs your code and never claims to have measured it.

Air-gapped, Local or Cloud

One switch decides where your code goes. Air-gapped and Local keep every call on your own hardware; Cloud sends calls to Claude on your own key, with the cost shown before the scan and a per-scan ceiling. An administrator can lock the cloud off.

Learning Requests

Ghost Architect™ learns your team's judgment. A reviewer records a verdict on a finding (a false positive, an accepted risk or the wrong severity) with a written reason and opens a Learning Request, much like a pull request for how findings are judged. A second person approves it and a third deploys it, and only then does anything change. Every step is logged and signed, any change rolls back, and older reports can be re-read under the new rules without the originals ever being altered.

Fewer false findings

Test, example, generated and documentation files are not reported as problems in your product, a finding must cite a file that is really in the code, and claims are checked against the code before they reach the report.

Corrected reports

After your team reviews a scan, download the original and the corrected report as PDFs, without scanning again.

Your whole fleet on one scan

Every machine and model in your cluster shares a scan's work, you choose which model does which job, and Claude can take only the calls your own machines are too busy for, up to the ceiling you set.

Ask about Ghost

Ask a question in plain words, or paste an error message, and Ghost answers from the manual and the error catalog that ship with your version, through your own model. When the manual does not cover the question, it says so instead of guessing.

The manual in the box

The full user's manual comes with the download, as a PDF and as Markdown, in docs/manual: getting started, scans and reports, the portal, the API, profiles, learning, and an appendix for every error message.

1. Get the tarball

Get the beta

Fill this in and the download link is emailed to you straight away. It works once and expires after 72 hours; come back here for another if you need one. The GitHub username is in case you would rather have access to the delivery repository. The model question is so we know the hardware the beta is running on.

Do you run your own model today?

Already have access to the delivery repository? gh release download v8.0.58 --repo EJWisner/ghost-local-delivery

Verify it before you unpack it

The file is about to become a global command on your machine. Check it first.

shasum -a 256 ghost-architect-local-8.0.58.tgz

On Windows PowerShell:

Get-FileHash -Algorithm SHA256 ghost-architect-local-8.0.58.tgz

Expected for 8.0.58:

4582fb9b97d44c92b85f1fec0e2b179adcff3551c84c7dc03671819a262d05cc

If the hash does not match, stop. Do not install it. Email [email protected] with what you got.

2. What you need

macOS 13 or later, Linux, or Windows 10 or later

The installer checks the OS floor and stops below it. Linux installs run as a service account with a data root the installer creates, and the Linux path is verified on a fresh machine for every release. Windows uses a PowerShell installer from an elevated prompt and registers Windows Services. Ghost Architect™ Local runs on real Windows hardware here, as two Windows Services, and the full test suite runs on that same machine on every push. What we have not done is watch a customer install from scratch on a machine that is not ours, so send us the console output either way.

Node.js 24 or later

The installer checks it and does not install it. Get it from nodejs.org or your version manager first.

A model server with a model pulled

Ollama, vLLM, LM Studio, Jan.ai or any OpenAI-compatible endpoint on your network. Our estate runs qwen3:14b on Ollama. The installer asks which and checks it answers.

A second machine, when you want the estate

Any of the three can be a spoke. Enrol it from Admin Console on the hub and it runs scans and reports them back; the model calls go to the hub's model server over your LAN.

3. Install

Everything happens inside the directory you unpack. Keep that directory; the global ghost command links into it rather than copying it.

tar -xzf ghost-architect-local-8.0.58.tgz
cd package
bash install/setup.sh              # macOS
sudo bash install/setup-linux.sh   # Linux
.\install\setup.ps1                 # Windows, elevated PowerShell

The installer changes four things on your machine, and INSTALL.md, in the directory you just unpacked, is the authoritative list of them. Re-running the installer is the supported upgrade path: your service account, data root and signing key are left alone.

4. First run and the trial

  1. Run ghost. The evaluation trial starts on first use, registered against this install, and the product tells you the day it ends.
  2. Point it at a codebase. Start with Recon: one planner call that sizes the codebase and tells you what a full scan would read, how many passes, and how long, with the basis of that estimate printed beside it.
  3. Open the portal it prints (https://localhost:3443 by default): the Fleet page, the queue, LLM Status, the reports. Trust its certificate the way the page tells you to.
  4. Enrol a second machine from Admin Console when you want the estate view. Every machine reports back to the hub.
  5. To keep it, convert from inside the product during the trial. The first month is half price and the licence arrives by email, signed, and activates against this install.

Zero egress, if that is how you work

Activation is the one call home (the licence key and an install id, nothing from your code). After it, turn on zero-egress mode and the product refuses every outbound destination at the code level and prints the inventory that proves it. Details on the security page.

Questions before you install?

[email protected] answers Local questions, or use the support chat with the platform set to Local. The walkthrough shows every page before you commit a machine to it.