ShellForge — Forge Local AI Agents

Forge Local AI Agents.
Governed.

Govern any AI coding agent — local or cloud. One policy across Goose, Claude Code, Copilot, Codex, and Gemini. Mac with local models. Linux server with API drivers. Your rules, everywhere.

Open Source · MIT License · macOS · Linux · Any GPU
⭐ Star on GitHub Quick Start →
$ brew tap AgentGuardHQ/tap
$ brew install shellforge
$ shellforge setup
── Step 1/6: Ollama ──
✓ Ollama running
✓ Model ready: qwen3:8b
── Step 2/6: Governance ──
✓ agentguard.yaml created (enforce, 3 policies)
── Step 3/6: Dagu ──
✓ Dagu installed
Setup Complete

$ shellforge agent "find test gaps in this repo"
[governance] loaded — enforce, 3 policies
[intent] list_files via json_block
[governance] allow | list_files
[intent] read_file via json_block
[governance] allow | read_file
[intent] run_shell via json_block
[governance] DENY | rm -rf / (destructive)
[correction] sending feedback to model...
[intent] run_shell via json_block
[governance] allow | go test ./...
done — 4 turns, 4 tool calls, 1 denial corrected

$ shellforge run claude "review open PRs"
[governed] Claude Code with AgentGuard hooks

Every Agent. Every Action. Governed.

Local models on your Mac. API drivers on your server. One governance layer for all of it.

🦙

Any Model, Any Provider

Local via Ollama (qwen3, mistral, phi4) on Mac/GPU Linux. Or use Claude, Copilot, Codex, Gemini APIs on headless servers. ShellForge governs them all.

🛡️

AgentGuard Governance

Policy-as-code in agentguard.yaml. Block force-push, restrict file writes, enforce timeouts. Monitor mode → enforce mode when ready.

⚡

Single Binary. No Daemons.

One Go binary, zero dependencies. Each agent command runs and exits. Cron-safe. No background processes, no distributed systems, no complexity.

🧠

Memory-Optimized

RTK strips 70-90% of terminal noise before the LLM sees it. KV cache quantization doubles agent capacity on 8GB Macs.

🔌

Multi-Driver Governance

Govern any CLI agent — Goose (local), Claude Code, Copilot CLI, Codex, Gemini. One governance policy across all drivers.

🔄

Dagu Orchestration

shellforge swarm — YAML DAG workflows with web dashboard. Schedule agents, define dependencies, run a 24/7 swarm on your Mac or Linux box.

Where AgentGuard Lives

Between every tool call and your environment. The gatekeeper.

INPUT 📝 Your Prompt Natural language task or command
↓
COMPRESS ⚡ RTK Rust Token Killer — strips 70-90% of noise
↓
QUANTIZE 🧠 TurboQuant 6x KV cache compression for large models
↓
INFER 🦙 Ollama Local inference · qwen3 · mistral · any GGUF
↓
EXECUTE 💻 Goose — AI Coding Agent Native Ollama support — reads, writes, edits, runs commands
↓
GOVERN 🛡️ AgentGuard — The Gateway allow / deny / audit every action before it reaches your machine
↓
EXECUTE 🌍 Your Environment Filesystem · Shell · Git · Network · APIs
↓
SCAN 🐾 DefenseClaw Cisco supply chain security — scan skills, verify MCP servers
↓
SANDBOX 🔒 OpenShell NVIDIA kernel isolation — Landlock + Seccomp

Governed Drivers

One governance policy. Any CLI agent. Every tool call checked.

Driver Command Model Status
Goose shellforge run aider Local (Ollama) governed
Claude Code shellforge run claude Claude API governed
Copilot CLI shellforge run copilot GitHub Copilot governed
Codex CLI shellforge run codex OpenAI coming soon
Gemini CLI shellforge run gemini Google coming soon
Built-in shellforge agent Local (Ollama) works now

The Full Stack

Seven technologies. One governed agent runtime. Here's how they fit together.

INPUT 📝 Your Prompt "Analyze this repo for test gaps"
↓
COMPRESS ⚡ RTK (Rust Token Killer) Strips 70-90% of terminal noise before the LLM sees it
↓
MODEL 🦙 Ollama Local inference — qwen3, mistral, any GGUF model
↓
EXECUTE 💻 Goose (Block) AI agent with native Ollama support — actually executes shell, files, git
↓ agent wants to: write file, run shell, push git
GOVERN 🛡️ AgentGuard — The Gateway Policy-as-code — allow/deny every action before it touches your environment
↓ approved actions only
EXECUTE 🌍 Your Environment Filesystem · Shell · Git · Network · APIs — the real world
↓
SCAN 🐾 DefenseClaw (Cisco, RSA 2026) Supply chain security — scan skills, verify MCP servers, AI BoM
↓
SANDBOX 🔒 OpenShell (NVIDIA, GTC 2026) Kernel-level isolation — Landlock + Seccomp, signed audit logs

ShellForge

Governed runtime
The control layer

Goose

AI coding agent
Executes the work

AgentGuard

Governance kernel
Enforces policy

Ollama

Local LLM inference
Powers the model

RTK

Token compression
70-90% fewer tokens

Dagu

Workflow orchestration
Schedules the swarm

OpenShell

Kernel sandbox
Isolates the process

Goose executes. AgentGuard governs. Ollama infers.
RTK compresses. Dagu orchestrates. OpenShell sandboxes.

All open source. All local. All on your Mac.

Star the repo to follow along.

Get Started in 60 Seconds

Install. Setup. Forge.

$ brew tap AgentGuardHQ/tap
$ brew install shellforge
$ shellforge setup
$ shellforge agent "analyze this repo for test gaps"


View on GitHub AgentGuard Core →