███╗   ███╗ ██████╗ ██████╗   █████╗ ██╗
  ████╗ ████║██╔═══██╗██╔══██╗ ██╔══██╗██║
  ██╔████╔██║██║   ██║██║  ██║ ███████║██║
  ██║╚██╔╝██║██║   ██║██║  ██║ ██╔══██║██║
  ██║ ╚═╝ ██║╚██████╔╝██████╔╝ ██║  ██║██║
  ╚═╝     ╚═╝ ╚═════╝ ╚═════╝  ╚═╝  ╚═╝╚═╝
  LOCAL AGENT ORCHESTRATION

COMMAND CENTER ↑ ↓ select · Enter open · q back

❯ Start a new task The specialist team plans, debates, executes, and verifies Select working directory Resume a task Model and parameters Orchestration profile Command screen System status Cloud model Türkçe / English Exit

MODEL mod-agent:latest · num_ctx 16384 FOLDER /Users/bg TEAM up to 24 logical agents · one model · internet on CLOUD off (local default)

mod-agent

mod-agent is a local AI model designed for coding, reasoning, tool use, and multi-agent orchestration on Apple Silicon.

Built on Qwen3.5 9B and optimized around the MLX / NVFP4 runtime, mod-agent is designed to work especially well with modAI CLI — a lightweight local agent orchestration harness for coordinating multiple logical AI agents through a single local model.

The idea is simple:

Run a capable multi-agent development environment locally without loading a separate large model for every agent.

mod-agent is intended for developers, researchers, automation workflows, and anyone experimenting with private on-device agent systems.

---

Quick Start

Run the model with Ollama:

ollama run studiobrn/mod-agent

16K variant:

ollama run studiobrn/mod-agent:16k

---

modAI CLI

mod-agent is designed to work with modAI, a terminal-based local multi-agent orchestration environment.

GitHub:

https://github.com/WeAreTheArtMakers/modAI

The CLI provides a command-center interface for managing agents, tasks, working directories, model parameters, context length, reasoning modes, and local/cloud model routing.

A typical architecture looks like this:

                    modAI CLI
                       │
                ┌──────┴──────┐
                │ Orchestrator │
                └──────┬──────┘
                       │
       ┌───────────────┼───────────────┐
       │               │               │
    Planner          Coder          Reviewer
       │               │               │
       ├───────────────┼───────────────┤
       │               │               │
   Researcher        Tester         Debugger
       │               │               │
       └───────────────┬───────────────┘
                       │
                 Request Queue
                       │
                       ▼
              studiobrn/mod-agent
                       │
                       ▼
                  Apple Silicon

Multiple logical agents can share the same inference backend instead of loading multiple copies of a large model into memory.

This makes agent orchestration practical even on machines with limited unified memory.

---

Capabilities

mod-agent is designed around agentic workloads including:

• Coding and repository analysis
• Tool calling
• Multi-step reasoning
• Planning
• Code review
• Debugging
• Testing
• Structured output
• Vision input
• Multi-agent coordination
• Terminal-driven automation
• Local-first execution
• Private on-device inference

---

Context Length

The model supports large context windows, while the actual runtime context can be adjusted depending on available memory and workload.

The modAI CLI allows num_ctx to be configured directly.

Recommended configurations:

8192     Lightweight tasks and simple agents
16384    General coding and agent workflows
24576    Larger repositories and reasoning tasks
32768    Extended coding and long-running agent workflows

For example:

num_ctx = 32768

A larger context window allows agents to retain more information from:

system instructions
tool definitions
conversation history
source files
terminal output
test results
error messages
patches
git diffs
previous agent actions

Higher context sizes also require more unified memory.

For systems with limited memory, 16K provides a good balance between context capacity and runtime efficiency.

For larger-memory Apple Silicon systems, 24K or 32K can be useful for longer coding sessions.

---

Agentic Coding Workflow

mod-agent is designed to work with repository-aware agents that inspect only the information required for the current task.

A typical workflow:

Inspect repository
        ↓
Identify relevant files
        ↓
Read required source
        ↓
Plan changes
        ↓
Edit code
        ↓
Run tests
        ↓
Inspect errors
        ↓
Review changes
        ↓
Finalize

Instead of blindly loading an entire repository into the context window, agents can progressively inspect the project and request the files or information they need.

This makes better use of limited context and unified memory.

---

Multi-Agent Orchestration

modAI can expose multiple specialized logical agents while using a single underlying local model.

Example roles may include:

Orchestrator
Planner
Coder
Reviewer
Researcher
Tester
Debugger
Documentation Agent
Security Reviewer
Architecture Agent

The orchestration layer manages task delegation, context, tools, and communication between agents.

The model remains shared.

Conceptually:

24 logical agents
        │
        ▼
orchestration layer
        │
        ▼
request queue
        │
        ▼
one local model
        │
        ▼
Apple Silicon

This approach reduces unnecessary memory consumption and makes local multi-agent systems significantly more practical.

---

Apple Silicon

mod-agent is designed primarily for Apple Silicon systems.

The MLX / NVFP4 model format is well suited to Apple’s unified memory architecture and allows the model to run locally without requiring a discrete NVIDIA GPU.

Suitable systems include:

MacBook Air
MacBook Pro
Mac mini
Mac Studio
Mac Pro with Apple Silicon

Memory requirements depend heavily on context length, active applications, parallel requests, and the number of models loaded at the same time.

For machines with limited unified memory, using a single loaded model with multiple logical agents is recommended.

---

Model Variants

mod-agent:latest

General-purpose mod-agent release.

ollama run studiobrn/mod-agent

mod-agent:16k

Variant configured for a larger default runtime context.

ollama run studiobrn/mod-agent:16k

Runtime context can also be controlled from compatible clients such as modAI CLI.

---

Example Use Cases

Local coding assistant
Autonomous coding workflows
Repository exploration
Code generation
Code review
Debugging
Test generation
Architecture planning
Documentation
Research agents
Tool-using assistants
Terminal automation
Structured reasoning
Multi-agent experiments
Private local AI systems

---

Design Philosophy

mod-agent follows a simple principle:

One capable local model can power many specialized logical agents.

Instead of assigning a complete model instance to every role, the orchestration layer manages specialization.

The agents provide different responsibilities.

The model provides the intelligence.

The harness provides coordination.

Agent Roles
    +
Shared Context
    +
Tool Access
    +
Task Delegation
    +
One Local Model
    =
Local Multi-Agent System

This architecture makes sophisticated agent workflows possible on consumer hardware while keeping inference local and under the user’s control.

---

modAI Project

modAI — Local Agent Orchestration

https://github.com/WeAreTheArtMakers/modAI

The project is focused on building lightweight infrastructure for local AI agents, coding workflows, orchestration, experimentation, and human-controlled automation.

Contributions and experimentation are welcome.

---

Ideas & Collaboration

We are interested in:

new agent roles
new local models
coding agents
agent orchestration strategies
developer tools
Apple Silicon optimization
local AI applications
tool integrations
experimental interfaces
automation concepts
research ideas

If you have an idea for a model, application, agent workflow, integration, or collaboration, feel free to get in touch.

Contact: studiobrn@gmail.com

---

License

mod-agent configuration, system instructions, orchestration additions, and original project components are released under the Apache License 2.0 where applicable.

The underlying model and all third-party components remain subject to their original licenses, copyright notices, attribution requirements, and applicable terms.

---

mod-agent

Local intelligence. Multiple agents. One model.