xNAUT runs your AI coding agents and records what they did. Every command, timestamped and chained into a SHA‑256 Merkle tree, exported as a report your client can verify — without sending a byte to the cloud.
brew install --cask 48nauts-operator/tap/xnaut
Free & open source · MIT · latest release · Apple Silicon, Intel & Windows x64 · macOS 13+
| Claude Code | 3 h 12 m |
| Codex | 0 h 51 m |
| Docker | 1 h 04 m |
| cargo · tests | 0 h 47 m |
| Total billable | 5 h 54 m |
THE APP
The Observatory lists every agent and every sandbox currently running, with the command to attach to any of them. Plan consumption sits across the top, because agents cost money and a number you cannot see is a number that surprises you.
01 — NAUTFLOW
An idea becomes a shipped change through fifteen stages and eight roles. xNAUT holds all of it — the documents, the tickets, the agents doing the work, and a Validator that refuses to let a build start on a broken chain of reasoning.
Analyst
Idea · Concept · Business case
PM · Architect · Security
Product requirements · Architecture · Data model · API design · Security review
Planner · PM
Development plan · Sprint stories · Executable tickets
Builder · Reviewer
Build · Test and review · Release · Engram learning
The Validator sits between Plan and Deliver. It reads the whole document chain and checks contract coverage, stage completeness, cross-stage consistency, testability and build readiness. A FAIL blocks the build — and it writes an acceptance script that runs against the thing you actually built, not against its own opinion of it.
02 — FEATURES
Every feature exists because running Claude Code, Codex, and friends in a plain terminal means losing track of who is doing what, where.
Launch claude, codex, gemini or any CLI agent from a user-editable registry. Live status pills — working, done, blocked — with real-time hook pushes. Click a pill, jump to its tab.
One modal to spin up isolated worktrees with a recipe where push and status both just behave — then drop any agent straight into the new tree.
Side-by-side diffs where agents leave inline annotations — summary, rationale, confidence — anchored to the exact hunk. Notes update live as agents write them.
Each project is a workspace with its own tabs, terminals and zellij session. Switch projects, get your exact context back — nothing bleeds between them.
Split any pane with a sandboxed browser, a markdown editor with Mermaid, a file tree or a code editor. Docs and previews live beside the shell, not in another app.
Twelve built-in themes with full ANSI palettes — Catppuccin, Tokyo Night, Nord, Dracula — plus a generator: describe a vibe, get a matching theme for the whole app.
Draw agent workflows as nodes — triggers, agents, approval gates, bounded retries — with versioning, validation, cost limits and model governance. Or just describe it: the LoopBuilder agent drafts the loop and repairs validator feedback itself.
An Obsidian-style vault inside the terminal — wikilinks, backlinks, tags, search and templates for work and personal notes. The Vault Librarian agent searches, files and writes notes on request.
Your vault's wikilinks and your codebase's imports, rendered as a force-directed orb — 2D or a slow-rotating 3D nebula. Switch between notes and code with one selector, all scanned locally in Rust.
A built-in MCP server on localhost turns the terminal into a tool your agents call. Claude Code, Codex or any MCP client can list projects, create and update tickets with conflict-safe revisions, and read or write project docs in the vault. Per-session bearer token, copy-ready connection object in Project Settings.
{
"xnaut": {
"type": "http",
"url": "http://127.0.0.1:<port>/mcp",
"headers": { "Authorization": "Bearer <session-token>" }
}
}
// tools: projects · tickets · vault docs — all local
03 — PROOF OF WORK
The Work Session Logger records every command with timestamps and duration, chains them into a SHA-256 Merkle tree, and exports a client-ready HTML report. A QR code on the report recomputes the root, so a report that has been edited stops matching the record it came from.
04 — LOCAL-FIRST AI
Explain Screen, AI chat and theme generation run against Ollama or LM Studio on localhost by default. Cloud providers are there when you want them — behind a privacy monitor that flags leaked keys and PII before a prompt leaves the app.
05 — HOW IT STACKS UP
Everyone's open source now. Nobody else fits a terminal, an agent fleet, and a notary into a 21 MB download.
| xNAUT | Warp | Cursor | Ghostty | |
|---|---|---|---|---|
| Price | Free · MIT | $20/mo credits | $20/mo | Free · MIT |
| Ships as | 21 MB native app | app + cloud | VS Code fork | native binary |
| Full power, no account | ✓ | AI needs account | — | ✓ |
| AI fully on localhost | ✓ Ollama · LM Studio | cloud credits | cloud models | no AI |
| Any CLI agent, live status | ✓ | its own agents | its own agents | — |
| Visual agent-loop designer | ✓ nodes + gates | cloud (Oz) | — | — |
| MCP server for agents | ✓ docs · tickets | — | client only | — |
| Hash-chained proof of billable work | ✓ Merkle + QR | — | — | — |
| Prompt privacy guard | ✓ keys · PII | — | — | — |
Ghostty is a great terminal. Warp is a great subscription. Cursor is a great IDE.
xNAUT is the whole bridge crew — in a 21 MB download, for free.
Snapshot July 2026. Warp: AGPL client, $20/mo Build plan for AI credits. Cursor: proprietary, $20/mo Pro; 3.x runs parallel/cloud agents — its own, not yours. Ghostty: MIT, no AI by design.
06 — FAQ
Everything you need to know about xNAUT, your agents, and the ∅ funding model.
The workspace where agent work happens and gets recorded. One 21 MB native app holds the terminals your agents run in, their live status, git worktrees, diffs with AI annotations, project docs, the plan they are working through — and a hash-chained log of every hour of it. Instead of four separate tools and a spreadsheet for the invoice.
MIT-licensed, no paid tier, no account, no trial that expires. The catch is there is no catch — we're backed by ∅ Nobody, so there's no investor to monetize you for. If you use cloud AI providers, you pay them directly; local models cost nothing.
Only what you point somewhere: prompts go to the AI provider you configure (localhost Ollama/LM Studio by default), and the app checks GitHub Releases for updates. The app has no telemetry and no analytics — there is no backend of ours to send anything to. (This website uses self-hosted, cookieless privacy-respecting analytics; the app sends nothing.) The privacy monitor even flags API keys and PII before a prompt leaves the app.
Any CLI agent. Claude Code, Codex, Gemini, Grok, OpenCode come preconfigured; anything else drops into the registry at ~/.config/xnaut/agents.toml. Live status pills work for all of them.
No. Explain Screen, chat and theme generation run against Ollama or LM Studio on localhost out of the box. Cloud providers (Anthropic, OpenAI, OpenRouter, Perplexity) are optional and use your own keys, stored locally.
Rust + Tauri v2. The UI renders in your OS's built-in webview instead of shipping a bundled Chromium, so the whole app is a ~21 MB download where Electron apps ship hundreds of megabytes.
Every command is recorded with timestamp and duration and chained into a SHA-256 Merkle tree. The exported report (HTML with a QR verification code) recomputes the root, so a report that has been edited stops matching its record. It proves the report matches the log it came from, not the quality of the work. The chain is computed on your own machine with no outside signature over it, so it is not evidence against whoever holds the log.
They solve a different problem. Warp is a terminal you subscribe to; Cursor is an editor. xNAUT is where agent work gets planned, run and recorded — it replaces Warp's terminal without the account, the credits or the cloud, and it sits beside Cursor rather than against it: Cursor for editing, xNAUT for running the fleet and proving the hours. Neither of them can hand you a signed record of what the agents did. See the comparison above for specifics.
macOS 13 Ventura or newer (Apple Silicon + Intel) is the primary platform; the app declares that floor and older Macs refuse to launch it. Windows x64 installers (.msi / .exe) ship with every release — treat them as early: the core terminal works, but workspace features that rely on zellij are macOS-first. No Linux builds yet.
Yes — that's the MCP server. Any MCP client can list projects, create and update tickets (revision-safe), and read or write project docs in the vault, all over localhost with a per-session token.
Fast — twenty-five releases so far, the latest on 2 August 2026. xNAUT tells you in-app when a new version lands, and everything that changed is on the releases page.
$ brew install --cask 48nauts-operator/tap/xnaut