XE Local AI Engine

XE Local AI Engine runs AI workloads locally on one machine. A single ASP.NET Core process serves the React management UI, exposes loopback-only endpoints under /api/local/v1 plus SignalR hubs, persists to SQLite with per-column encryption, and supervises the llama-server, sd-server, and whisper-server child processes that do the inference. The current source version is 1.0.0-rc.2, composed in eng/ReleaseVersion.props.

Just want to install and use the app? Start with the User Guide — download, install (Windows & Linux), first run, troubleshooting, and privacy, all in plain language. App downloads are on the Releases page.

Official binaries are portable-only and self-updating: a Velopack Portable.zip on Windows and an AppImage on Linux. They are unsigned, so verify CHECKSUMS.sha256 before running them.

Installing on behalf of an AI agent? An external agent can install, set up, start and connect to this engine without a browser. Set XE_ADMIN_EMAIL and XE_ADMIN_PASSWORD first, and save the one-time XE_MCP_KEY= and XE_RECOVERY_CODE= values it prints. Details: the Agentic Support install guide.

curl -fsSL https://raw.githubusercontent.com/w0rldx/XE-Local-AI-Engine.Source/main/install.sh | \
  bash -s -- --setup --start --install-skill
# PowerShell: set XE_ADMIN_EMAIL/XE_ADMIN_PASSWORD plus XE_SETUP=1, XE_START=1, XE_INSTALL_SKILL=1, then:
# irm https://raw.githubusercontent.com/w0rldx/XE-Local-AI-Engine.Source/main/install.ps1 | iex

What ships from this repo

Area	Features
Core	Chat · agents and work sessions · scheduler · Model Advisor · knowledge base · image generation and editing · audio transcription
Developer	Development Mode · Dev Workflows · Graph Workflows · compute tools · training · benchmarks · custom tools · MCP client and server (Agentic Support) · web search and fetch
Platform	desktop shell · update channels · vault key custody · External Apps
Runtimes	llama.cpp by default, Ollama opt-in, stable-diffusion.cpp, whisper.cpp, or a cloud provider. See Local runtime & providers.

Feature switches are node settings, edited under Settings → Node Settings → General → Features and listed in wiki 08. Dev Workflows, compute tools, custom tools and web access ship off, and the External Apps catalog ships empty. The project layout is in wiki 02.

Agentic Support

A same-machine external agent can install, configure, start and operate the node through install.sh / install.ps1 and an inbound MCP server at /api/local/v1/mcp/server. An agentic MCP key is operator-equivalent only for its enumerated MCP tools: it grants no Operator role or JWT and no arbitrary REST access. The listener stays loopback-only. See the install guide, the MCP client runbook, the shipped external-agent skill and ADR 0006.

Architecture rules

• The node opens no outbound control-plane connection; every egress traces to a feature an operator turned on.
• Cloud-provider credentials and external endpoint tokens stay local and must not be returned to the browser or written to logs/transcripts.
• Local admin endpoints must be loopback/local-only, authenticated, strict about Host/Origin, and secret-redacted.
• Installers and packaging must not create background autostart behavior unless explicitly opted in via --autostart (user-scope only); autostart is never the default.

Documentation map

• User Guide — install, first run, troubleshooting and privacy, for non-developers.
• Developer Wiki — the code-grounded architecture reference. Start at Home.md.
• AGENTS.md — contributor rules, repository map, and the authoritative validation gate set; CONTRIBUTING.md — how to propose a change and what a PR must state.
• Architecture Decision Records — design decisions and their code-level scope.
• AI runtime developer notes — narrow AI-seam maintenance rules.
• Troubleshooting — operator and developer page: data directory layout, logging switches, support bundle, unattended starts; end users start at the FAQ.
• Audit index — the dated audits and reviews kept in this repository.
• Technical/Security Architecture Dossier — a baseline-scoped external review of the implementation as of 2026-07-28. Its baseline hash belongs to the pre-consolidation history and does not resolve here, so read it as a dated snapshot, not a commit to check out.
• Component notes: Node Web Server and React Client.

Local development

Prerequisites: the .NET SDK pinned in global.json; Node.js matching XE-Local-AI-Engine.Client.React/package.json; pnpm; Python 3 with uv; the Aspire CLI; and on Linux/WSL setsid. A GPU is optional and Docker is not required by default — the app self-provisions llama.cpp and GGUF models on first run.

scripts/dev-start.sh     # start the isolated Aspire AppHost for this checkout
scripts/dev-status.sh    # resource states and endpoint URLs (--json); the port changes on every restart
scripts/dev-stop.sh      # the only sanctioned stop path — never `aspire stop --all`

These wrappers keep parallel checkouts from killing each other's instances; the cleanup contract is in scripts/README-dev-stop.md. dev-start.sh also mints and reuses a per-checkout, never-tracked XE-Local-AI-Engine.AppHost/.data/node.key, so encrypted dev data stays readable. If it mints a key next to data written under a different secret, it says so and names what to delete.

scripts/run-backend-tests.sh              # the backend gate: one Release build, every enrolled test project

cd XE-Local-AI-Engine.Client.React
pnpm run acceptance                       # validate + coverage thresholds + tooling tests + production bundle

AGENTS.md §Validation is authoritative for the full gate set, including analyzer requirements, pnpm run openapi:check after a backend contract change, the Python scope, and the opt-in live runners.

Publishing and releases

Linux ships a self-contained AppImage and Windows a framework-dependent Portable.zip that needs ASP.NET Core 10.0.12+. Both open the native desktop shell by default, without a console; the engine log is in <data dir>/logs. --debug keeps a console streaming that log, and CLI modes (--browser, --headless, --mcp-only, --status, --setup, …) keep theirs (wiki 11 §5). Run one instance per user-data directory — a second races on the SQLite database. The tag-triggered release.yml is the only release path.

Maintainer mechanics, artifact contracts and RC evidence live in publish/README.md; the update-channel story in docs/velopack-release-install-guide.md; the publication gate in docs/release-publication-checklist.md.

License and conduct

XE Local AI Engine is licensed under Apache-2.0. See LICENSE and NOTICE. Participation is governed by the Code of Conduct; contribution rules are in CONTRIBUTING.md.