Overview

Forge is a terminal-based autonomous AI coding agent that runs locally on the operator's machine. It connects to any OpenAI-compatible endpoint, to Anthropic via direct API key, to ChatGPT Codex via OAuth against an existing ChatGPT Plus subscription, or to local model servers such as LM Studio, Ollama, llama.cpp, and vLLM. No mandatory cloud, no telemetry. Claude subscription (Pro/Max) login is not supported — Anthropic's terms restrict subscription OAuth credentials to its own applications. Use an Anthropic API key instead.

The agent is built around a small Rust core and a terminal UI, and is designed to be readable and tailorable rather than feature-maximal. The codebase is intentionally small enough that an operator can read it, understand what the agent is doing, and modify it when their workflow demands something a vendor wouldn't ship.

A standalone graphical IDE built on the same Rust agent core is approaching public release as a second surface. Initial release ships on macOS as a signed .dmg; Linux and Windows follow. The CLI and the GUI run against the same agent and can be used independently — the GUI isn't a replacement for the terminal experience, and the terminal isn't a fallback for the GUI.

Capabilities

Project log

  • Forge IDE reaches v0.1.0 code stability

    The IDE is now stable enough to be used as the primary editor in production work on the maintainer's own flow. Reliability across all flows is not yet fully established — expected for a first release — but the core loop is solid. Distribution testing on the signed .dmg begins now; public release follows when that testing passes.

  • Forge v0.2.1 released

    Reliability release. Ends the edit_file "old_string not found" loop by returning bounded recovery hints (closest match, whitespace diffs, or a disambiguation list) instead of a bare error — especially valuable for weaker local models. Restores token counting on streaming OpenAI-compatible endpoints (mlx_lm, vLLM, llama.cpp, LM Studio, OpenAI), which had been silently breaking auto-compaction on long local sessions. Recovers tool calls that misbehaving servers leak as raw text so the agent no longer stalls at high context on mlx_lm. Installer now checks for curl, ripgrep, and bun up front. See the CHANGELOG.

  • Forge v0.2.0 released

    v0.2.0 ships streaming reasoning display for chain-of-thought models (compact "Thinking…" line with Ctrl+T to expand), fixes compatibility with strict OpenAI-compatible servers like mlx_lm that require the system message first, and corrects a Ctrl-key shortcut leak into the input field. It also formally completes the Claude subscription OAuth removal first patched into v0.1.0 — Anthropic is still fully supported via API key. See the CHANGELOG for details.

  • Forge v0.1.0 released

    Forge is now publicly available. Source is on GitHub at github.com/Vulkgryph/Forge under the Vulkgryph organization. Installation, configuration, and usage notes live in the repository README. Reach [email protected] for inquiries.

  • Project page expanded

    Forge's project page now describes the agent's capability surface ahead of the public release push. Repository and documentation links remain pending until the public push lands.

  • Project page established

    Forge now has a project page and Git-backed project log source. Detailed changelogs can remain in the repository when public.

Documents