PragmAI PragmAI
winget install --id=PragmAI.PragmAI -e Installs the auditable PragmAI command for setup, local diagnostics, verified updates, and restorative uninstallation.
winget install --id=PragmAI.PragmAI -e Installs the auditable PragmAI command for setup, local diagnostics, verified updates, and restorative uninstallation.
PragmAI Core is the auditable local component of PragmAI. It installs the Codex and Claude Code hooks, converts locally available technical metadata into a closed telemetry schema, removes content, and sends one aggregate event per exchange with measurable usage.
This repository exists so employees, customers, and auditors can verify exactly what leaves a workstation. It does not contain the PragmAI backend, dashboard, credentials, metric evaluation, business interpretation, recommendations, or the logic used to decide possible improvements.
PragmAI Core is not a substitute for professional, security, compliance, legal, financial, or operational advice. Its telemetry and any hosted estimates are informational and may be incomplete, unavailable, or inaccurate when a client does not expose the required data.
PragmAI Core may send an employee-authorized email plus technical counters, closed categories, and approved public model/plugin/skill labels. It never sends prompts, responses, transcripts, commands, arguments, tool output, file names, paths, URLs, session identifiers, arbitrary tool names, or custom plugin/skill names. Unknown names are emitted only as other.
Codex telemetry v5 can emit a closed numeric counterfactual for the current exchange. Telemetry v6 adds a deterministic sensitivity grid at the current compaction threshold, ±25,000 and ±50,000 tokens, plus the original limit. Telemetry v7 adds aggregate applied-code line counters and completed plugin calls. Telemetry v8 adds approved public labels for plugins and skills plus Claude Code active-time, edit-result, commit, pull-request, and subagent counters. Telemetry v9 can add a declared automation origin, a company-assigned numeric bot slot, and an installation-keyed HMAC of a scheduler run marker. No bot name or raw run identifier is sent. All source records are processed transiently; only aggregate counters and closed labels leave the workstation.
See the telemetry reference and a synthetic complete event.
On macOS, install the current release from the public Homebrew tap:
brew install mmamani93/tap/pragmai
Windows and Linux artifacts are available from the versioned GitHub Release. Windows releases include a per-user pragmai-windows-x64-setup.exe installer and retain the portable ZIP as a fallback. WinGet remains unavailable until Microsoft accepts the catalog submission; do not present it as an active installation method before then.
After installation, the public interface is:
pragmai setup
pragmai doctor
pragmai check-update
pragmai update
pragmai update-status
pragmai repair
pragmai uninstall
setup first detects Codex and Claude Code. It stops without changes if neither is installed, uses the only detected client automatically, or asks the employee to choose one or both when both are available. It then asks for the employee-authorized email and requests a short-lived pairing. It prints only a public code, which the employee authorizes from the manually shared one-use invitation link. The hosted service then delivers an individual revocable credential directly to the executable. The temporary link may travel through a trusted chat, but no permanent credential appears in chat, URLs, or command arguments.
The local connector stores that credential only in its private user configuration and binds every event to the invited company and email. doctor checks configuration, integrity, and package/hook synchronization without reading company analytics or printing credentials. After an authorized package or artifact update, repair atomically synchronizes the private copy used by hooks and retains a recoverable backup. uninstall removes only PragmAI-managed changes and restores the captured pre-installation configuration.
Windows standalone releases from 0.7.18 implement update: select the signed manifest version, try a WinGet download and fall back to the official GitHub release if the package/version or WinGet is absent. Verify the GitHub digest and executable version, replace through a separate helper, retain a backup, and run repair and doctor. A pending handoff is not success; wait for completion or use update-status. Older Windows executables need a one-time assisted artifact update. macOS and Linux continue using their package or official artifact.
The released executable includes its Python runtime; employees do not need Python installed. Python and PyInstaller are required only on a release machine:
python -m pip install -r requirements-build.txt
python -m unittest discover -s tests -p "test_*.py"
python scripts/build_standalone.py --clean
PyInstaller builds separately on each operating system. GitHub Actions produces macOS arm64, macOS x64, Windows x64, and Linux x64 artifacts from the same tagged source. The Windows job then uses NSIS 3.12.0 to build and exercise the installer and uninstaller without administrative privileges.
Public:
Private:
Generated copies of published connector versions are not kept on the default branch. Published source snapshots and binaries remain available through immutable Git tags and GitHub Releases; the default branch contains only the current auditable source and release workflow.
Use of PragmAI Core and the decision to install, defer, test, or apply an update are the user's responsibility. Users must determine whether the software is appropriate for their environment, obtain any required organizational authorization, verify release signatures and hashes, maintain backups, test compatibility, secure credentials and devices, and validate results before relying on them.
Update checks are informational and do not install software. Neither the availability of an update nor an update notice transfers responsibility for installation, configuration, compatibility, security, continuity, data protection, or recovery to PragmAI or its contributors. No maintenance, support, update, or notification obligation exists unless it is agreed separately in writing. See the supplemental disclaimer and MPL 2.0 Sections 6 and 7 in LICENSE.
PragmAI Core is licensed under the Mozilla Public License 2.0. The hosted PragmAI platform and private analysis components are not included in this license and may be governed by separate terms. The supplemental notice in LICENSE makes the user-responsibility, no-maintenance, warranty, and liability boundaries explicit without modifying the MPL 2.0 grant.