Gi

Gi is a Rust-first terminal coding-agent harness. It drives an LLM through a tool loop with explicit permissions and a workspace jail, and adds a few things on top of its base: interactive ask_user questions the model can ask mid-turn, expressive terminal themes, stronger project-instruction discovery, and richer SKILL.md capabilities. Local and OpenAI-compatible providers (Ollama, llama.cpp, vLLM, …) are first-class, not an afterthought.

The binary is gi. Config lives in ~/.gi/; environment variables use the GI_ prefix.

Gi is a fork of ultraworkers/claw-code (MIT). It keeps that workspace as its implementation base and rebrands it to gi. See the ownership disclaimer below.

What Gi adds

• ask_user runtime tool — the model can pause and ask a structured question (choices, free text) and continue the same turn with your answer. Non-TTY runs report interactive_required instead of blocking; timeouts are enforced without leaking reader threads.
• Terminal themes — gi-dark / gi-light, selected with GI_THEME or the /theme slash command and persisted to ~/.gi/settings.json. Reported by gi status --output-format json and gi doctor; JSON output stays ANSI-free.
• SKILL.md extensions — optional version, tags, required_tools, provider_hints, and references frontmatter, with missing-reference diagnostics surfaced in gi skills list --output-format json.
• Provider/model discovery — gi models (and a first-run prompt) scans your environment for configured providers (Anthropic, OpenAI, xAI, DashScope, Sakana, Kimi, GLM, Ollama), live-queries their model lists, and persists your pick.
• Operating principles baked into the system prompt: compare approaches, prefer the smallest verifiable change, stay provider-agnostic, and keep risky actions behind explicit approval.

The canonical roadmap is docs/gi-harness-plan.md.

Quick start

# 1. Clone and build
git clone https://github.com/danmaxis/gi
cd gi/rust
cargo build --workspace

# 2. Set your API key (an Anthropic API key — not a Claude subscription)
export ANTHROPIC_API_KEY="sk-ant-..."

# 3. Verify everything is wired up
./target/debug/gi doctor

# 4. Run a one-shot prompt
./target/debug/gi prompt "say hello"

# 5. Start an interactive session
./target/debug/gi

On Windows (PowerShell) the binary is gi.exe: use .\target\debug\gi.exe or cargo run -- prompt "say hello".

Local / OpenAI-compatible providers (Ollama, llama.cpp, vLLM) are supported via ANTHROPIC_BASE_URL / OLLAMA_HOST; see docs/local-openai-compatible-providers.md.

Repository shape

• rust/ — the Rust workspace and the gi CLI binary (crate gi-cli).
• USAGE.md — task-oriented usage guide (commands, auth, sessions, config, themes).
• docs/ — design notes; gi-harness-plan.md is the roadmap.
• PHILOSOPHY.md / ROADMAP.md / PARITY.md — intent, backlog, and Rust-port parity.

Verification

From the rust/ directory:

../scripts/fmt.sh --check
cargo build --workspace
cargo clippy --workspace
cargo test --workspace

Documentation map

• USAGE.md — commands, auth, sessions, config, terminal themes
• docs/gi-harness-plan.md — roadmap and design
• docs/local-openai-compatible-providers.md — Ollama / llama.cpp / vLLM setup
• docs/windows-install-release.md — Windows / PowerShell install notes
• rust/README.md — crate map and CLI surface
• CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md
• LICENSE — MIT

Ownership and affiliation

• Gi is a fork of ultraworkers/claw-code (MIT), which is itself part of the broader agent-code / Claude Code lineage. Upstream attributions are preserved.
• This repository does not claim ownership of the original source material and is not affiliated with, endorsed by, or maintained by Anthropic.
• Distributed under the MIT License.