I wanted Claude to delegate to a local model. That became MAF.

Description and use

MAF enables task delegation between a paid Claude model and a local model, coordinating locally via Taskwarrior, Git worktrees, and file inboxes, without external services.

Problem

Complex reasoning tasks required a paid Claude model, while implementation work could be handled by a local model, but existing tools did not support seamless delegation and coordination.

Outcome

MAF enables reliable delegation between paid and local models, reduces manual coordination, and supports mixed harnesses without requiring all team members to use the same provider.

Workflow

• 1. Install MAF via gem install maf. 2. Configure harnesses with maf add claude:architect opencode:backend-developer claude:reviewer. 3. Run agents interactively or in dispatch mode. 4. Agents read/write tasks on the Taskwarrior board and communicate via file inboxes. 5. Use maf guide to set up and customize roles or workflows.
• 5. Monitor coordination through the optional web dashboard (.maf/bin/dashboard).

Evidence

I started building MAF because I wanted to make AI more reliable and spend less on it.

Source

My idea was simple: use a paid claude based model for complex reasoning, planning, and judgment, then use a local model for implementation work.

Source

I started using Claude Code, but I still wanted to use a local model through Ollama and OpenCode for some side tasks.

Source

MAF is the glue that lets different coding-agent tools coordinate work on one project.

Source

A project-local Taskwarrior board records goals, tasks, and claims.

Source

You can use Claude Code with your subscription and OpenCode with a local model, or choose another combination of harnesses and models.

Source

MAF also installs Git hooks and harness‑specific files. Those are real changes to the local environment, even though the tool keeps its setup out of the clone's normal Git status.

Source