Skills

Forge: A Model-Agnostic AI Coding Harness and CLI — A Decision Guide Before Your First Use

Forge is best suited for model-agnostic AI coding tasks in the terminal, automatically routing each task to the optimal model by cost and capability. Getting started requires accepting the one-line shell install script (which also installs the desktop app by default — decline with n) and completing forge setup on first launch to connect models and security settings.

Official source · GitHub - adulari/forge

Install

curl -fsSL https://raw.githubusercontent.com/Adulari/forge/main/install.sh | sh

Who It's For

  • Developers who want to complete coding tasks in the terminal with a single native CLI/TUI, and want each task automatically routed to the model with the best cost × capability trade-off. It is a model-agnostic AI coding harness written in Rust, shipped as a single native binary with no Node, Python, or Bun runtime required.
  • If you want to use it via a desktop app, iOS/Android/Web companion apps, or an OpenAI-compatible local API, the official documentation states these forms are also provided.
  • Not a fit when: you need installation methods beyond a graphical setup wizard, or you depend on a specific plugin/extension ecosystem — the evidence does not provide these details, so this cannot be confirmed; which specific agents are supported and which language environments are supported are not stated in the official evidence and remain unknown — do not assume on your own.
  • Prerequisites: accessible model credentials. forge setup walks you through connecting models, security and cost settings, search, imports, and background services; API keys are stored in the OS keyring.

First Try in 3 Minutes

There is only one recommended main path: the official one-line shell install script.

  1. Prerequisite: a terminal that can run shell commands, with access to GitHub.
  2. Run in your terminal:
curl -fsSL https://raw.githubusercontent.com/Adulari/forge/main/install.sh | sh

The script downloads the matching release binary and installs it to ~/.local/bin.

  1. Side-effect checkpoint: the interactive install also installs the Forge desktop app by default. If you don't need it, answer n at the prompt (or set FORGE_DESKTOP=0 to skip). This is a pipe-to-shell remote script installation — verify that you trust the source before running it.
  2. After installation, launch forge. On first launch it automatically opens the forge setup wizard; follow the prompts to connect models and security/cost settings.
  3. For your first task, pick a low-risk analysis task, such as asking it to explain a piece of code or summarize the project structure in the current repository. Observable success signal: the task is dispatched and the model's answer/analysis is returned in the terminal UI. As for whether this process reads or writes your files, executes commands, or accesses the network, the evidence does not say — verify and gate this yourself before running any task that makes changes.

Other methods (secondary note only): in a cloned repository, the hooks, MCP servers, temper, and allow rules under .forge/ remain inactive until you run forge trust.

Use Cases

Case 1: Quickly understand an unfamiliar repository at low cost. Scenario: you've taken over an unfamiliar code project. Approach: launch forge's terminal UI in the repository directory and issue a read-only analysis task, such as "summarize this project's structure and entry points." It routes the task to an appropriate model via its routing mechanism, and you see the result directly in the TUI. Expected outcome: a project overview; this is an analysis task and does not modify files.

Case 2: One-shot (headless) execution of a single coding request. Scenario: scripted use, or you don't want to enter the interactive UI. Approach: invoke forge in one-shot/headless CLI mode and submit a clearly defined coding task (the evidence does not give the specific command-line argument format; what can be confirmed is only that this form exists). Expected outcome: the task is routed to the optimal model chosen by cost × capability and the output is returned. For coding tasks that modify files or execute commands, define the scope of changes before submitting — the evidence does not describe its side-effect behavior, so you must gate this at the environment level yourself.

Boundaries of Use

  • Installation side effects: the one-line install script installs the desktop app by default (interactive default is yes); decline with n or FORGE_DESKTOP=0.
  • Repository-level trust boundary: .forge/ hooks, MCP servers, and temper/allow rules in a cloned repository are inert by default and only take effect after an explicit forge trust — this is an officially stated security checkpoint.
  • Runtime environment: a single native Rust binary with no Node/Python/Bun/Electron dependencies. API keys are stored in the OS keyring.
  • Beyond the officially stated limitations, the following information is not provided in the evidence and remains unknown: which specific agent clients are supported, which natural/programming languages are covered, and whether review or diagnostic operations read/write files, execute commands, or access the network. Do not treat the unknown as "zero risk"; for any task that modifies files, executes commands, or calls external services, confirm the side effects before running it.