A Rust TUI coding agent. Connects to any OpenAI-compatible API (Ollama, OpenAI, Anthropic, Google Gemini, OpenRouter, etc.), your ChatGPT subscription via OpenAI Codex, or your GitHub Copilot subscription, and provides an interactive terminal interface for code generation, analysis, and assistance.

Prebuilt binaries:

Download from GitHub Releases.

From source:

cargo install --git https://github.com/leonardcser/smelt.git

Running smelt with no config file will launch an interactive setup wizard that walks you through selecting a provider and model.

With Ollama (local):

ollama pull qwen3.5:0.8b
smelt --model qwen3.5:0.8b --api-base http://localhost:11434/v1

With OpenAI:

read -s OPENAI_API_KEY && export OPENAI_API_KEY
smelt --model gpt-5.4 --api-base https://api.openai.com/v1 --api-key-env OPENAI_API_KEY

With OpenAI Codex (ChatGPT Pro/Plus subscription):

smelt auth          # log in with your ChatGPT account
smelt --model gpt-5.4    # use any Codex-supported model

With GitHub Copilot:

smelt auth                          # pick "GitHub Copilot", follow device-code prompt
smelt --model claude-sonnet-4.5     # use any model your Copilot plan exposes

With Anthropic:

read -s ANTHROPIC_API_KEY && export ANTHROPIC_API_KEY
smelt --model claude-opus-4-5 --api-base https://api.anthropic.com/v1 --api-key-env ANTHROPIC_API_KEY

• Tool use — file read/write/edit, glob, grep, bash, notebooks, web fetch/search
• Permission system — granular allow/ask/deny per tool, bash pattern, URL, and workspace-scoped approvals
• 4 modes — Normal, Plan, Apply, Yolo (Shift+Tab to cycle)
• Vim mode — full vi keybindings for the input editor
• Sessions — auto-save, resume, fork, rewind conversations
• Compaction — LLM-powered summarization to stay within context limits (auto-trigger threshold configurable via SMELT_COMPACT_THRESHOLD_PERCENT, default 80)
• Reasoning effort — configurable thinking depth (off/low/medium/high/max)
• File references — attach files with @path syntax
• Multi-agent — parallel subagents with inter-agent messaging (opt-in)
• Skills — on-demand specialized knowledge via SKILL.md files
• MCP — connect external tool servers via the Model Context Protocol
• Custom commands — user-defined slash commands via markdown files
• Custom instructions — project-level AGENTS.md files
• Input prediction — ghost text suggesting your next message
• Image support — paste from clipboard or reference image files
• Headless mode — scriptable, no TUI
• Interactive setup — guided first-run wizard and smelt auth for managing providers

Config file: ~/.config/smelt/config.yaml (respects $XDG_CONFIG_HOME).

providers:
  - name: ollama
    type: openai-compatible # or: openai, anthropic, codex, copilot
    api_base: http://localhost:11434/v1
    models:
      - qwen3.5:27b

  - name: openai
    type: openai
    api_base: https://api.openai.com/v1
    api_key_env: OPENAI_API_KEY
    models:
      - gpt-5.4

  - name: codex
    type: codex # uses ChatGPT subscription — models fetched automatically
    api_base: https://chatgpt.com/backend-api/codex

  - name: copilot
    type: copilot # uses GitHub Copilot subscription — models fetched automatically
    api_base: https://api.individual.githubcopilot.com

defaults:
  model: ollama/qwen3.5:27b # provider_name/model_name

auxiliary:
  model: openai/gpt-5.4-mini
  use_for:
    btw: false

settings:
  vim_mode: false
  auto_compact: false
  redact_secrets: true # on by default — scrubs secrets from user input and tool results before they reach the LLM

See the full documentation for all config options, CLI flags, keybindings, permissions, and more.

Full docs are available at leonardcser.github.io/smelt and can be built locally with Zensical:

uv tool install zensical
cd docs && zensical serve

cargo build       # compile
cargo run         # run
cargo test        # run tests
cargo fmt         # format
cargo clippy      # lint

Inspired by Claude Code.

Contributions welcome! Open an issue or pull request.

MIT — see LICENSE.