Jarvis

An AI coding assistant CLI, inspired by Iron Man's J.A.R.V.I.S.

Jarvis is an interactive REPL harness similar to Claude Code, built with Zig and libvaxis.

What works today

• Interactive REPL with streaming LLM responses (stdin readline loop)
• Anthropic provider — SSE streaming via anthropic-sdk-zig
• Bash tool — shell execution with dangerous-command blocking
• Multi-turn conversation history (Session)
• .env config — API key, base URL, model ID loaded from environment or .env file

Coming soon

• Full libvaxis TUI interface
• OpenAI-compatible streaming
• File tools (read / write / edit)
• Glob / Grep tools
• TOML config file (~/.config/jarvis/config.toml)

Quick Start

Prerequisites: Zig 0.15.2+, and anthropic-sdk-zig checked out as a sibling directory:

parent/
  jarvis/
  anthropic-sdk-zig/

# Build
zig build

# Run
zig build run

# Test
zig build test

Configuration

Copy .env.example to .env and fill in your values:

cp .env.example .env

ANTHROPIC_API_KEY=sk-ant-...
ANTHROPIC_BASE_URL=https://api.anthropic.com
MODEL_ID=claude-opus-4-5

Environment variables take precedence over .env values.

Architecture

src/
  main.zig          ← CLI entry point
  config.zig        ← Config loading (.env + env vars)
  core/             ← Session & token management
  llm/              ← Provider abstraction + Anthropic SDK wrapper
  tools/            ← Tool registry & implementations (bash)
  repl/             ← Agent loop & stdin REPL

Dependency direction (one-way, no cycles):

main → config
main → repl → llm → core
            → tools
            → core

Development

# Install pre-commit hooks
lefthook install

# Format check
zig fmt --check src/

# Run all tests
zig build test

Roadmap

• Anthropic SSE streaming
• Agent loop (tool call → execute → feed result back)
• Tool: bash execution (with dangerous-command guard)
• Multi-turn conversation history
• .env config loading
• Full libvaxis TUI loop
• OpenAI-compatible streaming
• Tool: file read / write / edit
• Tool: glob / grep
• TOML config file