FRAMEWORK

LegionForge Framework

Local-first AI agent framework with security in the foundation

built on LangGraph, runs on your hardware, enforces guardrails in deterministic code

What it is

LegionForge is the flagship project of the LegionForge ecosystem — a framework for building AI agent systems that run on your hardware and behave predictably under adversarial inputs.

It's built on LangGraph for the agent graph runtime, Ollama for local LLM hosting, and a custom security pipeline that wraps every tool call, every LLM invocation, and every input/output boundary in deterministic checks.

What's in the box

Local-first LangGraph runtime

Agents execute on your hardware. Cloud APIs are optional, not required. Switch providers via env var.

Deterministic security pipeline

Prompt-injection detection, tool revocation, capability-boundary, Ed25519 signing, adaptive threat rules — all in code, none in the model.

Three loop-protection layers

Step counter, action-history hash, token budget. A single failure can't loop forever.

Multi-provider LLM factory

Ollama, OpenAI, Anthropic, InceptionLabs behind one interface. Rate-limited, cost-estimated, cloud-fallback ready.

PostgreSQL state layer

Async pool with role separation, LangGraph checkpoint resumption, pgvector RAG, SHA-256 hash-chained audit log.

Gateway API

FastAPI with task queue, SSE streaming, A2A and MCP endpoints, Bearer auth, web UI with HITL approval gate.

Connectors

Discord, Telegram, Slack, WhatsApp, generic webhook. Drop-in chat-to-agent bridges.

Guardian sidecar

7-check deterministic pipeline on every tool call, running in its own container. Sub-5ms latency.

What makes it different

Compared to LegionForge Their model
Cloud agent platforms (OpenAI Operator, Anthropic Computer Use, Google Mariner) Runs on your hardware. Your data never leaves. Runs on their hardware. Your prompts and tool outputs pass through opaque infrastructure.
OSS agent frameworks (LangChain, AutoGen, CrewAI) Security enforced deterministically on every tool call. No opt-out. Flexible substrate. Guardrails are libraries you may or may not add.
Productivity agents (Devin, OpenDevin, OpenClaw) Auditable security stack. PostgreSQL audit chain. Role separation. Vary widely. Recent OpenClaw analysis surfaced 512 vulnerabilities, 8 critical.

Status

v0.7.1 Current version (alpha)
2247 Smoke tests passing
38 Integration tests
381 Operator tools

Get started

# Install
git clone https://github.com/LegionForge/LegionForge.git
cd LegionForge
pyenv install 3.11.15 && pyenv local 3.11.15
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Configure
export AGENT_HARDWARE_PROFILE=mac_m4_mini_16gb

# Bring up infrastructure and run smoke tests
make check
make db-init
make start
make test-smoke

Full setup walkthrough in the Getting Started guide.

Read more