Skip to content

LegionForge Documentation

A local-first, security-native AI agent platform. Security is enforced in the execution path — not layered on afterward.

This is the central documentation hub for the LegionForge ecosystem: the framework, the Guardian security sidecar, and the tools and apps built on top.

Get started with the framework Guardian quickstart LegionForge on GitHub

What's in the ecosystem

Core

Project What it is
LegionForge The framework. Local-first AI agent platform built on LangGraph, with a full security stack baked into every layer.
Guardian Deterministic security sidecar for AI agent frameworks. Drop-in protection against prompt injection, tool poisoning, and capability abuse.

Tools

Project What it is
llm-valet Cross-platform LLM lifecycle manager — auto-pause/resume Ollama based on resource pressure and gaming detection.
mcp-probe Connectivity and configuration advisor for MCP services you own or operate.
headroom System stability monitor — memory pressure, paging, AI-powered diagnostics.
hermes-tool-test-suite pytest harness for validating Hermes AI agent tool-calling reliability.
dev-rig Shared Python CI workflows, pre-commit config, and audit harness.

Apps

Project What it is
Jeli LegionForge's sovereign, cryptographically-attested personal memory framework.
ADHD-OS Personal assistant framework for those with ADHD.

Design principles

The whole LegionForge ecosystem is built around a small set of non-negotiable principles:

  • Fail-safe tiering — halt → sandbox/retry → degrade. Never silently succeed.
  • Human gates on all mutations — destructive actions cross a human-in-the-loop boundary.
  • Replace AI with determinism wherever possible — the LLM is the last resort, not the first.
  • Validate at trust boundaries, not at processing nodes — sanitize once, at the edge.
  • Privilege tied to tasks, not persistent to agents — capability-scoped, expires when the task ends.

Read more in Security Model.

Where to go next