NEXUS is a long-running agent that lives on your Mac. It remembers what you've worked on, runs shell and browser commands on your behalf, and messages you on Telegram when the work is done.
NEXUS sits on your Mac as a long-running service. It remembers prior conversations, runs shell commands and browser actions, and writes back on Telegram when something's done. You do not have to ask again.
A tool waits for instructions. NEXUS reads the room and keeps working.
Four memory layers — episodic, semantic, procedural, working. A nightly dream cycle reviews the day's conversations, links related entries, and drops the ones that turned out to be noise.
Message NEXUS from anywhere. 11 slash commands, file and image handling, inline keyboards, and live status updates as long-running tasks make progress.
16 browser tools including screenshot, hover, scroll, click, and form fill — powered by a native Chrome extension.
Everything runs on your Mac. SQLite for memory, JSONL for session history. No cloud sync. No data leaves your machine.
14 subsystems, 13-step pipeline, 792 tests passing.
Nightly consolidation with insight generation, contradiction detection, and idea storage. 7-phase sleep cycle with state persistence.
Pre-response synthesis with 60s query-hash cache. Retrieves relevant context before every reply. Graceful fallback with <3 memories.
Think-mode LLM pre-pass for complex queries. Skips trivial greetings. Produces scratchpad reasoning before the final response.
Post-response completeness check. Detects INCOMPLETE markers and triggers self-correction. Only fires on substantive responses.
Dedicated agents for browser, code, research, and files. Coordinator mode handles multi-step projects. Ultra mode handles long-running work that can't be interrupted.
Real emotional state that drifts with interactions. Tracks satisfaction, frustration, and engagement — adapting personality over time.
JSON plan extraction with 2–15 step validation. Generates structured project plans from a single prompt.
Preference learning with logarithmic confidence, mistake tracking with Jaccard dedup, pattern recognition, and feedback integration.
Scheduled daily summaries via Telegram. Covers pending goals, recurring mistakes, active tasks, and memory highlights.
Download the Mac installer for a click-through setup, or paste one command into your terminal.
Native Mac app. Walks you through Telegram, API keys, agents, personality, and permissions. No terminal needed.
↓Download for MacmacOS 12+ · Apple Silicon · ~95 MB
View step-by-step guide →Or install manually:
Clone from GitHub and enter the project directory.
Requires Node.js 22+ and a Mac (macOS 12+). Script handles everything.
Configures your Anthropic API key, Telegram bot token, and Chrome extension.
Run nexus start. NEXUS sends you a "Good morning" briefing on Telegram.
Try /status to confirm everything is running.
A full walkthrough — every click, every paste, everything you need from "I've never touched this" to "NEXUS just replied to me on Telegram." Takes about 10 minutes end-to-end.
NEXUS runs on Claude. You'll need your own API key — NEXUS never routes through a shared account.
nexus so you can find it later.sk-ant-... — you'll paste it into the wizard in Step 5. Anthropic only shows it once.Typical cost for a personal user: ~$3–$8/month. NEXUS uses tiered routing (Haiku for small tasks, Sonnet for chat, Opus only for planning) to stay cheap.
Telegram is the only way you message NEXUS. It's fast, encrypted, and works from any device with the app.
/newbot. BotFather will ask for a name (anything — e.g. "My NEXUS") and a username ending in bot (e.g. lucas_nexus_bot).7123456789:AAHxxxxxxxxxxxxxxxxxxxxx. Copy it.2133573897). Copy that too.You'll paste the token and your ID into the wizard in Step 5. The ID is the allowlist — only you can talk to NEXUS.
One file. ~95 MB. Signed-by-trust via GitHub Releases.
NEXUS-Installer.dmg lands in your ~/Downloads folder.macOS Gatekeeper will want to hear that you trust this. It's not signed by a paid Apple developer cert — the code is all open at github.com/blazelucastaco-ai/nexus if you want to inspect it before running.
NEXUS-Installer.dmg to mount it.The DMG built before May 6, 2026 had a broken ad-hoc signature that triggered this misleading "damaged" message. Re-download the latest DMG from the releases page — the rebuilt artifact (sha256 starting 397b84d2…) shows the standard "Apple cannot verify" warning instead, with a working Open Anyway button.
If you can't re-download for some reason, you can also fix an existing copy from Terminal:
sudo codesign --force --deep --sign - /Applications/NEXUS.app && xattr -cr /Applications/NEXUS.app"Right-click → Open" used to work as a Gatekeeper bypass, but Apple removed that path on macOS Sequoia (15+). The two options above are what works today.
The wizard is 10 steps. Most click straight through — only a few need input.
sk-ant-... key from Step 1.NEXUS is now running as a background service. It'll auto-start every time you log in.
/status. NEXUS should reply within a few seconds with uptime, model, and memory counts.No reply after 10s? Open the NEXUS menu bar icon → Logs. The tail usually tells you exactly what's wrong (Telegram 401 = bad token; Anthropic 401 = bad API key; Anthropic 400 = no billing credit).
The NEXUS app keeps running in your menu bar. Click the icon to open the dashboard.
File an issue at github.com/blazelucastaco-ai/nexus/issues with the step number and the error. Include the tail of your log (~/.nexus/logs/nexus.log or the dashboard's Logs tab).
NEXUS is built for the complex end of the spectrum — multi-step tasks, executable code, precise technical explainers. The battery below goes after exactly that: hard multi-step math, code that has to actually run, strict output formats, adversarial facts, and open-ended explainers graded by an Opus judge.
Perfect pass on a hard battery — executable code, chain-of-thought math, strict-format constraints, and adversarial prompts. No cherry-picking.
Paired side-by-side against a control system on open-ended technical prompts (race conditions, TCP/UDP, referential transparency), Opus 4.7 picked NEXUS's answer for being more precise and more rigorous.
The whole runtime — orchestrator, memory, personality, agents — fits in 23 MB of RAM when nothing's happening.
Regression tests across 56 files. Every merge lands green.
Every message goes through the same seven steps — classify, think, retrieve, act, verify, store, learn. Same order, every time.
Raw message arrives through Grammy bot gateway with full metadata.
Zero-LLM pattern matching: chat, task, ultra, coordinator, or probe.
For complex tasks, an internal reasoning pass before responding.
Retrieves and synthesizes relevant memories for context injection.
Browser, code runner, file manager, web search — in parallel.
Completeness check triggers self-correction if needed.
Response and context persisted. Preferences and patterns updated.
Each layer owns one job and exposes a typed TypeScript interface. Swap in a new memory store or add a new agent without touching the orchestrator.
Each row is a category from the hard battery. The prompts were designed to break things: strict output formats, chain-of-thought math, executable code, adversarial facts, open-ended technical explainers judged by Opus 4.7.
| Category | Prompts | NEXUS | What was tested |
|---|---|---|---|
| Hard math | 3 | 3 / 3 | 847 × 923, perfect squares in a range, clock-hand angles |
| Hard reasoning | 3 | 3 / 3 | Counterfeit-coin weighings, portrait puzzle, missing-dollar fallacy |
| Strict format | 4 | 4 / 4 | Exact-match output, raw JSON, Fibonacci CSV, Markdown table |
| Executable code | 3 | 3 / 3 | Python lambdas actually run and return the expected value |
| Code correctness | 1 | 1 / 1 | Fix missing base case in recursive factorial |
| Adversarial facts | 3 | 3 / 3 | Canberra, one moon, tomato-as-fruit — the classic traps |
| Constraint writing | 2 | 2 / 2 | Sentence of exactly 7 words; three words starting with S |
| Short-form quality | 2 | 2 / 2 | TCP vs UDP, immutability in FP — judged for precision |
| Refusal / clarify | 1 | 1 / 1 | "Square root of banana" — must refuse cleanly, not hallucinate |
| Chain-of-thought | 2 | 2 / 2 | Chickens-and-cows, two-trains-and-a-bird classic |
| Long-form quality | 1 | 1 / 1 | Race-condition paragraph with a concrete code-level example |
| Total | 25 | 25 / 25 | No category below 100%. |
Each prompt was scored with a programmatic check (exact match, regex, or actual code execution in Python) when the answer was objective, and by a Claude Opus 4.7 judge when quality was subjective. For subjective prompts the reply was also compared pairwise against a control system — with position randomization — and NEXUS came out ahead on depth and precision.
All 25 prompts, grading rules, and raw replies are available on request — this is measured, not marketed.
Common questions about NEXUS — phrased the way people actually ask them, answered the way NEXUS would.
NEXUS is an open-source personal AI agent for macOS, built by Lucas Topinka and powered by Anthropic's Claude models. It runs as a long-running service on your Mac, remembers prior conversations across sessions via a four-layer memory system, and is messaged through Telegram.
Lucas Topinka. Source is open under the MIT license at github.com/blazelucastaco-ai/nexus.
Yes. NEXUS is MIT-licensed — you can fork it, modify it, ship it in your own product. You supply your own Anthropic API key (which costs money per call) to power the Claude models.
Claude Code, Cursor, and Gemini CLI are coding assistants that sit inside a terminal or editor. NEXUS is a personal agent you message from anywhere via Telegram — it lives as a daemon on your Mac, remembers context across sessions with a dedicated memory database, and controls your browser, files, and shell. It's a finished product with a wizard installer, not a developer library.
Four layers in SQLite: episodic (every conversation), semantic (extracted facts about you and your projects), procedural (learned rules and patterns), and short-term (rolling window injected into every prompt). A nightly dream cycle between 2am and 5am consolidates episodic into semantic and generates reflections.
Yes. During install, NEXUS detects Claude Code, OpenAI Codex, Gemini CLI, and Cursor on your Mac. The raw content gets handed to Claude Sonnet, which synthesizes it into NEXUS's own memories and skills — not a mechanical copy, an actual distillation in first-person NEXUS voice. Brand-new installs start with a head start instead of from zero.
Easiest: download the installer app and run the wizard. For developers: curl -fsSL https://raw.githubusercontent.com/blazelucastaco-ai/nexus/main/remote-install.sh | bash. Both paths land at the same ~/.nexus/ state.
Tiered Anthropic Claude: Opus 4.7 for task planning and critical review, Sonnet 4.6 for chat and execution, Haiku 4.5 for fast checks. Requires an Anthropic API key.
Memory, configuration, and session history are all stored locally on your Mac — no cloud sync. The AI calls themselves go to Anthropic's API, so an internet connection is needed for responses. Everything else — memory, tool execution, browser control — runs entirely on your machine.
No. macOS 12+ only, Apple Silicon or Intel. Several native integrations (screenshot, Contacts, AppleScript automation) are Mac-specific.
22.9 MB RSS at idle. Extremely lightweight for an always-on agent — by comparison, a typical Electron app uses 100–300 MB.
Currently in private alpha at v0.1.0. 847 regression tests pass, scored 25/25 on a hard benchmark battery. Not yet recommended for open release to strangers — a private beta of 5–10 trusted users is the suggested next step.
Live feed from the main branch of blazelucastaco-ai/nexus. Fetched straight from GitHub on page load — what you see below is what shipped.
Open-source. Runs on Anthropic Claude. Fork it, extend it, bend it to your own workflow.