OpenRouter is a unified AI gateway supporting 200+ frontier and uncensored open-weight models under a single API key.

Step-by-Step Key Generation:

• Visit openrouter.ai/keys and log in with Google, GitHub, or email.
• Click Create Key, name it Ghost Writer, and set optional credit limits.
• Copy the key (begins with sk-or-v1-...).
• Your key is encrypted in your device's browser localStorage only and is never transmitted to any DIGITAL GHOST servers.

Click the model selector in the topbar to open the categorized Model Picker with instant filter chips:

• ⚡ Ultra-Fast: Gemini 2.0 Flash, Claude 3.5 Haiku, DeepSeek V3, Llama 3.3 70B Nitro. Outputs 150–250+ tokens/second.
• 👑 Flagships: Claude 3.5 Sonnet, GPT-4o, Gemini 2.5 Pro. Unmatched depth for full-stack systems and architecture.
• 🧠 Reasoning: DeepSeek R1, o1-mini. Native chain-of-thought verification for algorithms and math.
• 🛡️ Uncensored: Hermes 3 405B, Dolphin Mixtral, Grok 3. No guardrails, perfect for red teaming, CTFs, and payload analysis.

Ghost Writer features automatic OpenRouter Provider Dispatch to optimize latency and generation speed:

How it works:

• Highest Throughput Routing: Ghost Writer instructs OpenRouter to route your request to the data center hardware cluster currently streaming the highest tokens/sec.
• Nitro Capacity: Compatible models (like Llama 3.3 70B Nitro) bypass queuing and run on dedicated high-speed clusters.
• Auto-Fallbacks: If a provider suffers temporary degradation, OpenRouter automatically fails over to the next healthy replica.

Toggle or customize your routing priority anytime under Settings → General → Routing Priority.

Ghost Writer's persona is engineered for directness, technical accuracy, and zero preachiness:

• No Disclaimers: Ghost Writer will not lecture, preach, or recite boilerplate ethical disclaimers.
• Full Exploit PoCs & Auditing: Generates exploit proofs-of-concept, fuzzers, shellcode analysis, CTF solvers, and binary reverse engineering breakdowns.
• Best Models for Offensive SecOps: Pick Hermes 3 405B, Dolphin Mixtral, or Grok 3 for unfiltered security research.

Easily feed source code, configuration files, and data logs directly into Ghost Writer:

• Click the paperclip button (📎) or simply drag and drop files onto the chat interface.
• Supported file types: .py, .js, .ts, .json, .txt, .csv, .md, .html, .sh, .sql, .c, .rs, .go.
• Files are read instantly in your browser via client-side FileReader. No files are uploaded to third-party file hosting servers.
• Token estimates update automatically in the topbar context counter.

Ghost Writer specializes in single-file tools, scripts, and web apps. Test them instantly without copying them out!

• Every assistant code block with HTML/JS displays a ▶ Run Sandbox button.
• Clicking it immediately switches to the 🖥️ Sandbox tab and executes the code in an isolated container.
• Edit the HTML/CSS/JS live in the right editor panel and click Run to test edits.
• View live console output (`console.log`, `console.error`) in the built-in Sandbox Console.
• Export or download the ready-to-use HTML file with one click.

Switch your agent's persona, model, and sampling parameters with a single click from the Presets bar:

• 🏎️ Speed Coder: Gemini 2.0 Flash / Haiku — Instant script generation, automation one-liners, low temperature.
• 🧠 Deep Reasoning: DeepSeek R1 / o1-mini — Algorithmic problem solving, complex math, architecture planning.
• 🛡️ Offensive / Red Team: Hermes 3 405B / Dolphin Mixtral — Unfiltered vulnerability research, exploit PoCs, payload crafting.
• 🎯 Refactor & Review: Claude 3.5 Sonnet — Rigorous code reviews, security hardening, bug triage.

Ghost Writer is architected with strict operational security (OpSec) standards:

• Zero Logging: There is no backend database tracking your conversations, prompts, or API keys.
• Direct Browser-to-API Calls: Requests travel directly from your browser to the OpenRouter completion endpoints.
• Local Persistence: All chat histories and custom models live in browser localStorage.
• Full Backup & Wipe: Export your full encrypted backup or wipe all device data with one click under Settings → Data.

Instead of merely predicting next tokens, Ghost Writer operates as an autonomous agent in a ReAct loop:

1. THINK (Reasoning): The agent analyzes your query and formulates a plan using internal reasoning tokens.
2. ACT (Tool Calling): Dispatches structured function calls: web_search(...), document_retrieval(...), or read_webpage(...).
3. OBSERVE (Inspection): Executes the action and parses returned data, web pages, or vector snippets.
4. SYNTHESIZE & RESPOND: Constructs an authoritative, verified answer with grounded source citations.

Feed your AI custom architectural specifications, user manuals, and codebases via Settings → Docs & Knowledge:

• Multi-Format Drag & Drop: Upload .pdf, .txt, .md, .json, .csv, or code files. PDFs are parsed and extracted directly in your browser.
• Free Vector Embeddings: Choose between the built-in ⚡ Client Vector Engine (100% free, runs offline with zero latency) or the 🤗 Hugging Face Serverless Inference API (all-MiniLM-L6-v2).
• Semantic Chunking: Documents are automatically segmented into overlapping semantic chunks with normalized vector embeddings.
• Agentic Retrieval: When you ask about your systems, the agent autonomously calls document_retrieval(query) to find the most relevant chunks.

Get up-to-date facts, current news, releases, and external documentation beyond static LLM training cutoffs:

• Zero API Key Required: Powered by @purili/web-search with resilient fallbacks to DuckDuckGo and Wikipedia.
• Automatic Decision: The agent detects when current info is needed, plans a search query, and executes the search.
• Webpage Reading: Uses read_webpage(url) to inspect full articles or technical documentation from returned results.
• Source Grounding: Clean hyperlinks and citation lists are appended to the response for verification.

Eliminate manual debugging cycles with real-time error reflection and autonomous code healing:

• Runtime Exception Trapping: Captures uncaught exceptions, promise rejections, syntax errors, and Python tracebacks with line-accurate telemetry.
• Autonomous Repair Pipeline: The agent inspects the broken code, diagnoses the root cause, and synthesizes a hardened patch without user intervention.
• Single-Click Auto-Fix: When an error occurs in the sandbox, click Auto-Fix to let Ghost Writer rewrite the code and hot-reload the running sandbox.
• Chat Diagnostic Fallback: If API keys are paused, transfers the exact error trace into the chat composer with one tap.

Execute real Python scripts in your browser with zero server dependencies or containers:

• WebAssembly Sandboxing: Official Pyodide CPython 3.12 engine running safely client-side in WebAssembly.
• Standard Library & Math: Full support for math, json, random, re, datetime, algorithms, data analysis, and simulated security tooling.
• Live Stdio Stream: Standard output (sys.stdout) and errors (sys.stderr) stream directly into the live sandbox console.
• Graphical & HTML Rendering: Generates HTML reports, ASCII tables, and dynamic visual output in the sandbox frame.

Seamless bidirectional feedback bridge between the live visual sandbox and the reasoning agent:

• Context-Preserving Feedback: Click Feedback to Chat to bundle current artifact code, language runtime, and console logs into chat.
• Guided Refactoring Presets: Instantly launch one-click workflows for UI modernization, dark theme polish, red team audits, and architecture cleanup.
• Zero Copy-Paste Friction: Never manually copy code back and forth between browser tabs. Ghost Writer keeps state synchronized.

Ghost Writer supports Anthropic's open Model Context Protocol (MCP 2024-11-05) client-side without needing Supabase or external servers:

• Zero Server Infrastructure Needed: Connect directly to remote MCP servers via JSON-RPC 2.0 over HTTP POST or SSE stream.
• Dynamic Tool Discovery: Queries tools/list from any connected server, registers the tools into the agent's function-calling registry, and executes them over tools/call.
• Built-in Utilities: Comes pre-loaded with an in-memory MCP suite (advanced math calculator, international timezone converter, text analytics, UUID generation).
• Interactive Playground: Test any discovered MCP tool manually in Settings → MCP Servers before executing them via agent reasoning.