Jarvis

A polished terminal coding cockpit for @advaitlakhera/jarvis.

Jarvis is a workspace-safe coding agent with a fast TypeScript/Node.js CLI, a panel-style Ink TUI, durable local memory, explicit shell approvals, unified diff output, and a broad provider catalog inspired by jcode.

Install

npm install -g @advaitlakhera/jarvis
jarvis

Or run without installing:

npx @advaitlakhera/jarvis

The package requires Node.js 20 or newer. This repository can also be developed locally with:

npm install
npm run build
npm test

Start coding

jarvis                              # interactive TUI
jarvis -p "fix the failing tests"   # one-shot mode
jarvis -p "run the test suite" -y   # auto-approve shell commands
jarvis --workspace ~/src/my-app
jarvis --provider ollama --model qwen3-coder

The TUI has a compact status header, readable transcript panels, a live thinking state, inline shell approvals, bounded tool output, and keyboard shortcuts:

| Shortcut / command | Action | | --- | --- | | /help | Show all commands | | /providers | List every provider profile | | /models | List models for the active provider | | /provider openrouter | Switch provider and persist the choice | | /model openrouter/openai/gpt-4.1 | Switch model, including qualified ids | | /level lite\|full\|ultra\|off | Change the ponytail guidance level | | /remember name \| fact | Save a durable local project fact | | /memory | Inspect saved memory | | ctrl+l | Clear the transcript | | ctrl+c | Exit |

Providers and models

Jarvis includes profiles for:

OpenCode Zen, OpenCode Go, OpenRouter, OrcaRouter, DeepSeek, Z.ai, Kimi, Moonshot AI, 302.AI, Fireworks, MiniMax, Meta Muse, Alibaba Coding Plan, Google Gemini, Azure OpenAI, GitHub Copilot, LM Studio, Ollama, Nebius, Scaleway, STACKIT, Baseten, Cortecs, Hugging Face, Firmware, OpenAI/Codex, Claude, and custom OpenAI-compatible endpoints.

See the live catalog with:

jarvis providers
jarvis models openrouter

Profiles use the standard endpoint protocol and read API keys from environment variables or Replit Secrets. Jarvis never writes secrets into its config file. OAuth login flows are intentionally not faked; for subscription providers, provide a compatible token or use the provider's supported environment setup.

The default profile is OpenCode Zen:

export JARVIS_API_KEY="..."
export JARVIS_PROVIDER=opencode
export JARVIS_MODEL=hy3-free
jarvis

Other common examples:

export OPENAI_API_KEY="..."
jarvis --provider openai --model gpt-4.1

export ANTHROPIC_API_KEY="..."
jarvis --provider claude --model claude-sonnet-4

jarvis --provider ollama --model qwen3-coder
jarvis --provider lmstudio --model local-model

Additional settings:

JARVIS_REQUEST_TIMEOUT=120000
JARVIS_MAX_TOKENS=8192
JARVIS_CONFIG=~/.jarvis/config.json
JARVIS_MEMORY=~/.jarvis/memory.json
JARVIS_AZURE_ENDPOINT=https://resource.openai.azure.com/...

jarvis config prints the configuration overview. /provider and /model choices are saved in ~/.jarvis/config.json; API keys are never saved there.

Safety and tools

The agent can list files, read files, write files, make unique-anchor edits, run shell commands, and save memory. Every path is resolved against the selected workspace and rejected if it escapes. Shell commands are denied by default until approved in the TUI; use -y only for trusted, self-contained tasks.

Package publishing

The package is configured for public npm publishing under the requested scope:

npm run build
npm publish --access public

Publishing requires an npm login or automation token in the publishing environment. It is not performed automatically by the development workflow.

Existing Python import

The original Python/Textual implementation remains in jarvis/ and its offline test suite remains available as a migration reference. The root npm package and jarvis binary are now the primary implementation.