mrig Agent

Overview

A personal, local, model-agnostic agent for the terminal. It writes and changes software, runs commands, works with files and answers questions, using Claude through your API key or a model running on your own hardware. Swapping models means editing a config file, not code.

The problem

Coding agents tie you to one vendor's model and one vendor's idea of what may run on your machine. Anyone who wants to use a model on their own hardware, for privacy or cost, usually has to leave the agent they already know.

The approach

One agent loop, with the model behind a provider interface. Anthropic models and anything that speaks the OpenAI protocol (Ollama, LM Studio, llama.cpp, vLLM, OpenRouter) are configured in JSON, merged from a global file, a per-project file, environment variables and flags, in that order of precedence. Tools that only read run freely; anything that changes files or runs a command asks first, with an "always" answer that adds a rule for the session.

Outcome

• Ships as a single binary per platform: Linux, macOS (Apple Silicon and Intel) and Windows
• The agent loop is tested against a mock model, so the test suite needs no network
• Lint (Biome, zero warnings), typecheck, tests and build run as one command
• Sessions are saved as JSONL and resumable, and per-run token counts and cost are shown

What it does

• Any model, by config

  Built-in profiles for Claude, OpenAI and local servers, and a generic OpenAI-compatible kind for everything else, each with its own pricing so the cost line is honest.

• Permissions you can read

  Read, glob and grep never prompt. Edit, write and bash ask, and allow rules such as bash:git status or edit:src/** make the answer permanent. A yolo switch exists and is off by default.

• Project memory

  Reads AGENTS.md from the working directory up to the git root, plus a personal one, into the system prompt, so conventions travel with the repository.

• Headless mode for scripts

  mrig -p runs one prompt and exits with a meaningful code: done, error, stopped at the step limit, or aborted.