AgentAleph: Local Coding Agent and GGUF Manager on llama.cpp

Overview

AgentAleph is a local-first AI coding agent and GGUF model manager built as a Tauri desktop app, with inference powered by llama.cpp. It brings model discovery, downloads, loading, and project work into one application, so developers can run an agent against a local model without a hosted inference service, API keys, or an account.

The agent can inspect and edit files and run commands in a selected project folder, with per-tool permissions and a read-only plan mode. The documented bundled setup targets Linux x64; other platforms need their own llama-server binaries. Optional web searches, downloads, and update checks still use the network.

Key Features

• Browse and download GGUF models from Hugging Face, then load, unload, and manage local models.
• Run streaming chat and an agent tool loop through llama.cpp's OpenAI-compatible API.
• Use file-reading, search, editing, writing, web, and shell tools with allow, ask, or deny permissions.
• Choose build or read-only plan mode, and review file-edit previews with an undo option.
• Select native tool calling or GBNF-constrained tool calls, with automatic routing available.
• Add project context from AGENTS.md or CLAUDE.md, local skills, and attached files or text.
• Resume interrupted model downloads and estimate hardware fit from available GPU VRAM and system RAM.

Use Cases

• A developer who wants to try local-model-assisted coding without configuring a separate inference server.
• A privacy-conscious user who wants prompts processed on their own machine and can avoid the optional network features.
• A Linux user who wants one desktop interface for finding GGUF models and using them on project files.
• A developer who wants agent-assisted edits and shell commands to require explicit approval, or wants a read-only planning pass.

Getting Started

On Linux x64, install the documented prerequisites, then:

git clone https://github.com/matiasA/AgentAleph.git
cd AgentAleph
npm install
./scripts/setup-llama.sh
npm run tauri dev

See the README for production builds, CPU-only setup, GPU backend options, and model guidance.

License in Plain Words

No standard license was detected. Without a license, the default copyright rules apply: you may read the code, but not reuse it. Check the repository or ask the authors.

Alternatives

• codeg: Codeg coordinates coding-agent sessions across desktop, server, Docker, and mobile, while AgentAleph is a Linux app with local GGUF inference and approval-based tools.
• atlas: Atlas manages multiple coding agents and links their work to Git commits, while AgentAleph focuses on a local llama.cpp coding agent and GGUF model management.
• grok-build: Grok Build is a terminal coding agent with scripting and ACP support, while AgentAleph is a Linux desktop app centered on local GGUF models and approval-based tools.

Considerations

• This is a new beta project with recent commits and releases, but the measured contributor history is concentrated in one person. Its short history provides limited evidence about long-term maintenance.
• The documented packaged target is Linux x64. Windows and macOS require separate llama-server binaries, and the README does not describe bundled installers for those platforms.
• llama-server is downloaded by the setup script rather than committed, and the default flavor is Vulkan. CPU-only and other backend variants require selecting a flavor during setup.
• Shell commands run with the project's working directory but are not an OS sandbox. Permissions and plan mode reduce risk but do not isolate commands from the host system.
• Web searches, model downloads, update checks, and approved shell commands can use the network, so local inference does not mean every operation is offline.
• GitHub and Google connections are placeholders; the project also lacks multiple sessions, multiple loaded models, and vision or multimodal support.