Fusion

A Rust CLI tool for managing local Ollama and MLX runtimes on macOS.

Overview

Fusion is a Rust CLI that manages local Ollama and MLX runtimes for development. It handles service startup, shutdown, status reporting, and now includes a first-class prompt runner that talks to the managed HTTP APIs directly. All behaviour is driven by a persistent TOML configuration file rather than ad-hoc environment variables, making the tool predictable across shells and sessions.

Getting Started

# Install dependencies and build
cargo build

# Print CLI help
cargo run -- --help

# Start the managed runtimes
cargo run -- ollama up
cargo run -- mlx up

Binary installation

cargo install --path .
fusion ollama up
fusion mlx up

Configuration

Fusion stores all runtime settings in ~/.config/fusion/config.toml (or the platform-equivalent using dirs::config_dir()). The file is created on first use with sensible defaults and can be managed via the CLI:

fusion config show             # dump the current file
fusion config path             # print the path to config.toml
fusion config edit             # create symlink to edit
fusion config reset            # reset to default values

The configuration file contains sections for both services:

[ollama_server]
host = "127.0.0.1"
port = 11434

[ollama_run]
model = "llama3.2:3b"
system_prompt = "You are a helpful assistant."
temperature = 0.8
stream = false

[mlx_server]
host = "127.0.0.1"
port = 8080
model = "mlx-community/Llama-3.2-3B-Instruct-4bit"

[mlx_run]
model = "mlx-community/Llama-3.2-3B-Instruct-4bit"
system_prompt = "You are a helpful assistant."
temperature = 0.7
stream = false

Logs, PID files, and runtime state are stored under each service's directory in ~/.config/fusion/<service>/. Override the project root for tests by setting FUSION_PROJECT_ROOT; the config location can be redirected with FUSION_CONFIG_DIR.

CLI Usage

fusion ollama up
fusion ollama down [--force]
fusion ollama ps
fusion ollama log
fusion ollama run <prompt> [--model <name>] [--temperature <value>] [--system <prompt>]

fusion mlx up
fusion mlx down [--force]
fusion mlx ps
fusion mlx log
fusion mlx run <prompt> [--model <name>] [--temperature <value>] [--system <prompt>]

# global commands
fusion ps
fusion config <show|edit|path|reset>

The run subcommand issues an HTTP request to the managed runtime using the defaults from config.toml, merging any CLI overrides for the model, system prompt, or temperature. Both services speak the OpenAI-compatible /v1/chat/completions API, so the CLI sends identical payloads and reuses the same streaming logic regardless of backend. The config family offers read/write access without leaving the terminal.

Testing

The project mirrors the original testing culture:

• Core unit tests live next to modules in src/core/, covering configuration persistence, service construction, and process lifecycle helpers.
• Integration tests in tests/llm_commands.rs drive the CLI entry points with a mock process driver and lightweight HTTP stubs, ensuring command wiring, configuration updates, and prompt execution stay consistent without launching real runtimes.

Run the full suite with:

cargo fmt --all
cargo clippy --all-targets -- -D warnings
cargo test

Project Structure

• src/core/paths.rs – project root, PID directory, and config file resolution
• src/core/config.rs – strongly-typed TOML configuration management
• src/core/services.rs – ManagedService definitions plus config-driven loaders
• src/core/process.rs – PID/log helpers and pluggable process driver
• src/cli/commands/ – lifecycle and configuration command handlers for managed runtimes
• src/cli/run/ – shared OpenAI-compatible run pipeline reused by each managed service
• tests/service_lifecycle.rs – integration tests for service up/down/ps/log operations
• tests/run_commands.rs – integration tests for run command execution and payload validation
• tests/config_commands.rs – integration tests for configuration management