agent-foundry-local 0.1.10

Local-first AI agent platform with formal handoff protocol for regulated industries

Local Agent Foundry

Agents building agents. Offline. Auditable. Regulated-industry ready.

A desktop platform for running, managing, and orchestrating AI agents entirely offline. Agents build, test, and deploy other agents — with full audit trails and compliance documentation for NHS, finance, and government.

Why

No existing platform combines:

• Local-only operation (air-gapped, GDPR-compliant)
• Agent orchestration (multi-agent with formal handoffs)
• Observability (audit trail, cost tracking, health)
• Quality gates (behavioural tests for agents)
• Regulated industry readiness (compliance docs, model cards)

Quick Start

# Install
pip install agent-foundry-local

# Or from source
git clone https://github.com/vystartasv/agent-foundry.git
cd agent-foundry
pip install -e ".[dev]"

# Start an agent
foundry agent start builder --model qwen-8b --tools terminal,file

# Run a task
foundry agent run builder "Create a test file"

# Hand off to a reviewer
foundry agent start reviewer --model qwen-8b --tools terminal,file
foundry handoff create --from builder --to reviewer --task "Review the work"

# Check quality gates
foundry gate run --agent builder --file output.txt

# Export audit log
foundry audit export --output compliance-report.json

Architecture

Six layers, running entirely on your machine:

Desktop UI (Tauri) ·················· [Phase 2]
CLI (`foundry`) ····················· [MVP ✓]
Agent Engine · Handoff Protocol ★ · Quality Gates · Observability
Model Runtime (oMLX · llama.cpp · Ollama)
Tool Registry (MCP · Hermes skills)
Storage (SQLite)

The handoff protocol is the core differentiator — a formal standard for agents to pass work between each other with full state preservation and verification. Every handoff is auditable.

For Regulated Industries

Pre-built compliance documentation included in docs/compliance/:

• DPIA Template — Data Protection Impact Assessment
• Model Cards — Per-model capability and limitation documentation
• GDPR Checklist — Article-by-article compliance verification
• NHS DTAC Alignment — Digital Technology Assessment Criteria mapping
• ISO 27001 Control Mapping — All 71 Annex A controls mapped
• Deployment Architecture — On-prem deployment patterns, data flows

Zero cloud dependency. Air-gap capable. No data leaves your device.

Requirements

• Python 3.11+
• macOS (Apple Silicon), Linux, or Windows
• 16GB+ RAM (32GB recommended for 8B+ models)
• No network required (offline operation after model download)

Model Backends

Backend	Status	Notes
oMLX	✅ Supported	Apple Silicon, local inference
llama.cpp	✅ Supported	Cross-platform GGUF models
Ollama	✅ Supported	Easy model management
Dummy	✅ Testing	Simulated backend for tests

Commands

foundry agent start|stop|list|stats|run    Manage agents
foundry handoff create|accept|complete|list|show    Formal work transfer
foundry gate run                          Quality checks
foundry skill discover|list|show          Agent skills
foundry audit show|export|stats           Audit trail
foundry backend health|models|test        Backend management

Development

pip install -e ".[dev]"
pytest                    # 163 tests
foundry --help            # CLI reference

Documentation

• Architecture — Full system design
• Compliance Pack — Regulated industry procurement docs
• Getting Started — Setup and first agent workflow

Roadmap

• CLI MVP — agent engine, handoff, gates, audit
• Compliance pack — NHS, GDPR, ISO 27001
• M365 Bridge — SharePoint, Teams, Graph API integration
• Multi-machine orchestration
• Cloud model fallback (opt-in)
• Desktop UI (Tauri) — dashboard, logs, model management

License

MIT — see LICENSE