Metadata-Version: 2.4
Name: gulutux
Version: 0.4.1
Summary: Multi-provider agentic coding CLI — swap the brain (Nebius Token Factory, Cloudflare, Google, any OpenAI-compatible), keep the fox.
License: MIT
Project-URL: Homepage, https://github.com/autokeren/gulutux
Project-URL: Repository, https://github.com/autokeren/gulutux
Keywords: ai,agent,cli,coding,nemotron,nebius,llm
Requires-Python: >=3.8
Description-Content-Type: text/markdown

# 🦊 gulutux

**Multi-provider agentic coding CLI.** Swap the brain — Nebius Token Factory,
Cloudflare Workers AI, Google, or any OpenAI-compatible endpoint — keep the fox.

> Named after my childhood nickname. Gulutux writes the code, its ghost agents
> run in parallel, and Ghostfox — the stealth browser — sees the web.

## What it is

gulutux is a terminal-native agentic coding assistant built in Go:

- **Multi-provider model layer** — one agent loop, any brain: Nebius Token
  Factory (NVIDIA Nemotron), Cloudflare Workers AI, Google AI Studio, OpenAI,
  or any OpenAI-compatible endpoint (Ollama, vLLM, proxies).
- **Token Factory Sandboxes execution** — commands run in ephemeral Nebius
  microVM containers (`sandbox` tool: one-shot runs or named sessions), with
  automatic sandbox→local fallback.
- **Cost-aware routing** — per-model pricing table, live token/USD metering
  from real API usage, and a session budget guard that downshifts to a fast
  model when the budget is reached.
- **Ghost agents** — spawn parallel sub-agents (`spawn_agent`) with
  capability-scoped tool allowlists and structured result mailboxes.
- **Evidence-led proofs** — acceptance criteria, recorded verification, and a
  `SHIP / BLOCKED / NEEDS_HUMAN_REVIEW` verdict gated by human approval.
  Deploy tools are blocked at the registry level until a proof is approved for
  the exact git commit.
- **Real browsers at its side** — a generic Chrome (CDP) layer and **Ghostfox**,
  a stealth browser driven through its MCP server: navigate, snapshot, click,
  type, screenshot with anti-detection identities.
- **Tavily web search** built in for grounded, real-time research.
- **Checkpoints (time travel)**, per-project persistent memory, SQLite kanban,
  and a vibe-security guard over shell/file tools.

## Build & run

Requires Go 1.22+.

```bash
make build        # produces ./gx
./gx              # interactive TUI
./gx "explain this repo" --task    # one-shot task
./gx --proof replay examples/proof-demo/proof-run.json   # replay a proof, zero API keys
```

Configure providers in `~/.config/gulutux/config.yaml` (see
`config.example.yaml`). Environment: `GULUTUX_*` (e.g. `GULUTUX_API_KEY`).

**Live demo:** https://gulutux.autokeren.com — this landing page was built,
rendered, and proof-verified **by the agent itself** in the demo video.

## Status

**v0.1.x** — foundation + Nebius stack verified live. All 24 Go packages pass
vet + tests. Verified end-to-end with NVIDIA Nemotron on Nebius Token Factory:

- agent thinking with `nvidia/nemotron-3-super-120b-a12b` (with automatic
  downshift to `Nemotron-3_5-Lightning` when a session budget is exceeded)
- Token Factory Sandboxes client (Contree API): spawn instances, subprocesses
  inside running VMs, one-shot runs, named sessions
- live dogfood run: the agent researched landing-page design via Tavily,
  wrote a page, rendered it in the Ghostfox stealth browser, and issued a
  `SHIP` proof — 13 turns, $0.08 of model spend

## Origin & license

MIT © 2026 Ajat Sudrajat.

gulutux is the successor of my prior open-source project
[autokeren](https://github.com/autokeren/autokeren) (MIT) — a Cloudflare-first
agentic CLI. The Go engine core, tool suite, ghost sub-agent manager, and
evidence-led proof system were carried over and are being significantly
extended for this project: a rewritten multi-provider architecture, native
Nebius Token Factory integration (inference + sandboxes), Ghostfox browser
integration, and a structured agent state machine. The prior project is
disclosed here explicitly, in the spirit of honest open source.

Built during the **Nebius x NVIDIA Global AI Hackathon** — running on
Nebius AI Cloud, thinking with NVIDIA Nemotron.
