System Documentation

High-speed access to open-weights models. Ideal for coding and rapid inquiry.

• Groq Cloud: GPT-OSS, Llama 3.3, Qwen 3
• NVIDIA NIM: Qwen 3 Coder 480B, Gemma 4 31B, DeepSeek V3.2, Codestral, Devstral
• CLI: Linked-account login or no-Google direct mode, folder-only or machine-wide access, SSH profiles, remote file workspace tools, Playwright MCP, named agents, autopilot runs, local plan board, verified self-update, and local or Vault-backed provider keys
• ZeroThink Logic: Disabled

Full access to the Pentagon Protocol and proprietary closed-source models.

• ZeroThink Logic: Active (Reasoning Layer)
• xAI: Grok 4, Grok 4 Fast, Grok Code Fast 1
• OpenAI: GPT-5.4, GPT-5 Mini, GPT-5.1
• Google: Gemini 2.5 Pro / Flash / Flash-Lite / Gemini 3 previews
• Billing: Cancel any time. Monthly non-refundable. Yearly refund requests within 7 days via shaf@talktoai.org

The Oracle route is tuned for both crypto scenarios and broader strategic calls. Modes include crypto, general, launch, and risk.

Oracle now defaults to Groq + GPT-OSS 120B for a strong speed / reasoning balance, while still letting you switch provider and model directly from the desk. Llama 3.3 70B, Compound, and the NVIDIA frontier set are there when you want a different read.

The Oracle page includes a Solana support card so people who use the desk heavily can donate directly to the ZeroThink build.

If someone wants ZeroThink without Google login first, the lean browser route is exposed at /guest. That lane accepts local Groq or NVIDIA keys, keeps memory/history off, does not connect to the ZeroThink account/Vault, and stores the key only in the browser.

The ZeroThink CLI, aka Agent Zero, installs on Windows PowerShell, macOS, and Linux. It can either pair to the same ZeroThink account as the web app or run in a no-Google direct mode with a local Groq or NVIDIA key on the machine. It now includes a real first-run setup guide on plain zerothink, beginner help like zerothink tutorial, zerothink help quickstart, zerothink help sessions, zerothink help plan, zerothink help browser, zerothink help mcp, zerothink help remote, zerothink help agents, zerothink key status, zerothink status, zerothink runtime, zerothink models coding, in-terminal /guide, a local /plan board, local web search and page browse, a first-class Playwright MCP browser lane, smarter short-prefix slash handling like /he, slash-command autocomplete when the optional prompt-toolkit dependency is present, a real remote workspace layer for SSH, SCP, and remote file read/write/replace flows, plus an explicit autopilot layer with saved named workers like zero1, zero2, and zero3. In the interactive prompt, plain-English build/fix requests now route into the real tool stack so Agent Zero can actually create folders, write files, and carry out bounded coding tasks instead of only narrating commands, and you can switch between /steer auto, /steer plan, and /steer confirm if you want tasks staged before execution. Use /go to run a staged task and /stop to discard it or request stop for the latest saved agent. Slash commands like /help only work inside the interactive ZERO> prompt. The CLI now runs in clean operator mode by default, so the full ZeroThink reasoning trace stays in the web app and Studio instead of being dumped into terminal replies.

The CLI supports direct file injection, workspace scans, file search, file writes and replacements, zip creation and extraction, remote SCP transfers, remote file inspection and edits on saved server profiles, repo-level guidance files like ZEROTHINK.md, CLAUDE.md, AGENTS.md, and .claude/CLAUDE.md, local web search, saved browser pages, Playwright MCP automation for dynamic pages, page attach flows, and the same packed file strategy used in the web app, so large source files and config dumps are easier to analyze without immediately blowing the window. The CLI reasoning layer now follows a stronger private orient → inspect → plan → act → verify → report workflow instead of staying in generic assistant mode.

The public install and manual page is exposed at /cli. The CLI can run in a safer folder-only mode for one repo or switch to machine mode when you want shell and server operator access, and the guide now walks beginners through install, linked-account mode, no-Google direct mode, first prompt, help topics, model selection, autocomplete, local checkpoint/resume flows, local planning, and the limits of direct mode. It also has a built-in zerothink update command that verifies the published CLI hash before replacing the local script, zerothink status for a quick local readiness snapshot, a remote workspace lane via zerothink remote ..., a local MCP layer with a built-in Playwright browser profile via zerothink mcp add-playwright and zerothink playwright open https://example.com, plus self-driven agent workflows like zerothink autopilot "audit this repo", zerothink agent create zero1 "map the auth flow", zerothink agent log zero1, zerothink stop zero1, and in-prompt /stop. For coding, the current recommended defaults are openai/gpt-oss-120b on Groq and qwen/qwen3-coder-480b-a35b-instruct on NVIDIA, and the fastest way to inspect the curated shortlist is zerothink models coding. If you just type a normal task like make a new project folder for shafaet.com inside the interactive prompt, Agent Zero now routes that into execution mode automatically instead of staying in pure chat mode, unless you switch to a staging lane like /steer plan.

If you skip Google login in the CLI, ZeroThink can still auto-detect GROQ_API_KEY or NVIDIA_API_KEY from your local shell, or you can store them locally with zerothink key set groq / zerothink key set nvidia. That lane does not import Vault keys, memory, saved history, linked-account session state, or Pro account state. Local CLI browser tools like zerothink search, zerothink browse, /search-web, /page, and local Playwright MCP browser commands still work because they are machine-local terminal helpers rather than website account features. For Groq and NVIDIA, the CLI can also fall back to that local direct-key lane automatically if the linked provider lane fails. If you ever want the exact live route instead of asking the model to guess, use zerothink runtime or /runtime. If you already have the CLI installed, the safe update command is simply zerothink update.

Analyzes the pure logical constraints of the request.

Applies abstract association and probabilistic weighting.

Checks alignment with the 11:11 Math of Goodness.