A floating AI companion that turns any laptop into a sovereign workstation
with 480+ modules. Runs 100% offline. Sees no data. Tracks nothing. Ever.
🆕 April 2026 update: ai_imp now builds anything for you — describe what you want in plain English, watch a working app pop out in its own window seconds later. Production-grade SaaS with built-in compliance. Mobile apps. Trading companions that start in paper mode and only graduate to real money after they've proven themselves. All on your laptop. Whether you want to peek under the hood or not, our template gallery is now MIT-licensed on GitHub — fully open, free to use anywhere, even with Cursor, Bolt, or Claude Code.
HyperChat • The AI • Anti-Palantir Shield • 4-Model AI • Your Desktop • Privacy Promise • Pricing • Get Started
Your data leaves your machine. The AI is lobotomised. The interface is a sidebar stuck to the edge of your IDE. You pay a cloud bill that grows every month. The "assistant" stalls on sensitive topics and forgets what you told it last week. Your tools are scattered across 14 browser tabs, 3 apps, and a Discord server.
You wanted a command center. You got a subscription to someone else's surveillance system.
ImpForge is the alternative.
Forget app windows that eat your screen. ImpForge lives in HyperChat — a small, beautiful AI panel that floats at the edge of your display.
It's tiny. It's smart. It moves on its own.
Think of HyperChat as a portal. You tell it what you need, and the right tools appear on your desktop as floating windows. An IDE here. A spreadsheet there. A chart in the corner. You arrange them however you want.
HyperChat is how you command your entire desktop.
It shows you what your AI is thinking. It shows you your goals, deadlines, and predictions. It streams every LLM token, every tool call, every state change through a live event protocol so you see the AI working — no black box, no "loading spinner theatre". It lets you switch between 5 levels of AI autonomy — from "only answer when asked" to "handle everything yourself" — each level guarded by a cryptographic audit chain that logs every decision.
No other tool works like this. Most AI assistants are a sidebar. HyperChat is a command center.
ImpForge has a built-in design intelligence system that no other AI tool has:
Not "we saved your chat history." Real memory. ImpForge builds a picture of who you are over time — your skills, your preferences, your projects, your goals. Every conversation makes it smarter about you specifically.
Ask it something on Monday. Reference it on Friday. It connects the dots.
Most AIs just throw words at you. ImpForge's AI has an actual thought process. When you ask a hard question, it checks your code, looks at your past decisions, considers what worked before, and then answers.
You can watch it think through HyperChat's Thinking Panel — see which parts of its brain light up while it works.
Working on a React component? It becomes a frontend expert. Writing a business proposal? Switches to a business consultant. Doing your taxes? Becomes a German tax advisor who knows the current deadlines.
You don't configure this. It just happens.
For the big decisions, your AI assembles a virtual boardroom: a CTO, a CFO, a legal advisor, a security expert — 10 specialists who debate your question from different angles. They argue. They vote (weighted by a Hebbian trust score that learns which advisor has been right in your context). Then they give you a recommendation with dissenting opinions clearly noted.
Like having a team of advisors who never sleep and never charge by the hour.
Three operating tiers, one kill-switch. Supervised (the AI asks before every action). Guarded (auto-runs safe actions, asks for sensitive ones). Autonomous (full self-governance for trusted long-running jobs). Every decision at every tier is recorded in a SHA-256 Merkle audit chain — tamper-evident, mathematically verifiable, yours alone.
The more you work with ImpForge, the more it learns. After a week, it anticipates what you need. After a month, it feels like it reads your mind.
All that learning stays on your machine. Delete the app, it's gone. No traces anywhere.
"Your AI assistant that even Palantir can't spy on."
Data surveillance platforms like Palantir work by fusing data from multiple sources into searchable profiles. They need three things to exfiltrate data (the "Lethal Trifecta"): untrusted input, access to sensitive data, and the ability to communicate externally.
ImpForge breaks all three.
Choose your protection level during setup. It controls EVERYTHING — AI routing, storage, network access, telemetry, and sharing.
| Tier | Name | What It Means |
|---|---|---|
| 0 | Fortress | Air-gapped. Zero network. Military grade. Your data physically cannot leave your device. |
| 1 | Vault | Local-only processing. Encrypted storage. No cloud sync. Enterprise GDPR strict. |
| 2 | Shield | Local AI primary. Anonymous cloud fallback with differential privacy. No PII ever sent. |
| 3 | Bridge | Encrypted cloud sync. E2E encryption. P2P sharing with colleagues. |
| 4 | Open | Full cloud integration. Maximum productivity. Privacy still protected by encryption. |
While competitors scramble to comply with the August 2026 deadline, ImpForge ships compliant by default:
ImpForge can mathematically prove that your data was processed correctly — without revealing what the data is:
Every line of ImpForge code was written assuming someone hostile is watching. Because they are.
AI Chat. Code Editor. Document Writer. Spreadsheets. Email. Calendar. Notes. PDF Viewer. Docker Manager. GitHub Dashboard. Trading View. Browser. Terminal. CRM. Analytics. Presentation Builder. Campaign Manager. Multi-repo orchestrator. Live-broker dashboard (Alpaca, IBKR, Binance, Kraken, Coinbase). And 380 more.
Each one is a building block. Tell HyperChat to open it — it appears as a floating window on your desktop. Drag it anywhere. Resize it. Pin it. Make it transparent. Build the workstation you've always wanted.
ImpForge doesn't give you "sample projects." It gives you builders. Ask HyperChat for a SaaS, a mobile app, a desktop app, a CRM, an e-commerce store, a firmware project, or a single-page site — and the AI scaffolds the whole thing, runs it in an isolated sandbox, and iteratively improves it with quality feedback until it works. Twelve categories. One command each.
No iframe juggling. No "deploy to see it." When ai_imp finishes building your app, a fresh window pops out on your desktop running the live preview — hot-reloaded, interactive, yours to click through. Move it to a second monitor. Leave it open while you describe the next change. It updates as you speak.
Building a fintech app? ai_imp ships FINRA-aware audit logging by default. Healthcare? HIPAA-ready schema. SaaS platform? Multi-tenant isolation auditing catches the one-line bug that causes 92 % of SaaS data breaches — before your migration ever hits the database.
We put the scaffolding blueprints on GitHub under an MIT license so anyone — even people who'll never buy ImpForge — can read, fork, or redistribute them.
template.json v1 specificationUse them standalone with any tool you like. Or let ImpForge's engine turn them into live apps in one conversation. Your call, both work.
ai_imp can follow your investment strategies — but no agent gets real money until it has proven itself. Every trading companion starts in paper mode. It stays there until it's racked up 200 winning sessions, with zero regressions, in market conditions below the VIX breaker threshold. Only then does promotion to real-money micro-trading unlock. You never wake up to a surprise.
Create a "Coding" workspace: IDE on the left, terminal at the bottom, AI chat on the right. Create a "Trading" workspace: charts everywhere, news feed, portfolio tracker. Switch between them in one click.
Each workspace remembers everything — window positions, AI context, active tools, even which personality your AI was using.
It's like having multiple desktops on steroids. The kind of setup you've only seen in movies or on trading floors that cost EUR 10,000+. Except it's on your laptop, and the base version is free.
This isn't a section we added because lawyers told us to. This is why we built ImpForge.
Not analytics. Not usage data. Not error reports. Not "anonymized telemetry." Nothing. Zero. Null.
When you use ImpForge, no packet of data leaves your computer heading to our servers. We don't have servers that receive user data. We didn't build them. We don't want them.
Your conversations with the AI? Yours. Your documents? Yours. Your browsing habits? We don't even know them. There is no business model that involves your data. Our business model is: we make good software, you pay EUR 25/month if you want the pro features. That's it.
No cookies. No analytics scripts. No "we use cookies to improve your experience." No fingerprinting. No device IDs. The app doesn't phone home. Not on launch. Not ever.
ImpForge's AI doesn't send your questions to a cloud API. It runs locally, on your hardware, using open-source AI models. Your conversations never touch the internet.
Every database on your machine is encrypted with AES-256 — the same standard used by governments and militaries. Protected by a separate password that even the AI cannot enter. Even if someone steals your laptop, they get encrypted noise.
Using Cursor, Copilot, or Claude Code alongside ImpForge? They can read files on your disk — but ImpForge's databases look like random gibberish to them. Encrypted at rest. Always.
We didn't add privacy features after building the product. We built the product around privacy. The encryption, the local processing, the zero-collection policy — these aren't toggles you can turn off. They're how the software works at its core.
AiImp Technology's commitment: We will never collect, store, sell, share, or analyze any user data. Not today. Not when we're bigger. Not when investors ask us to. This is a founding principle, not a business decision.
| Free | Pro (EUR 25/mo) | Founder's Edition (EUR 750 one-time) | |
|---|---|---|---|
| HyperChat AI Panel | Yes | Yes | Yes |
| All 400+ modules | Yes | Yes | Yes |
| Local AI (runs on your PC) | Yes | Yes | Yes |
| Encrypted databases | Yes | Yes | Yes |
| AI personalities | 50 | 472 | 472 |
| Quality rules | 50K | 158K+ | 158K+ |
| Board of Directors | — | Yes | Yes |
| Digital memory | Basic | Full | Full |
| Enterprise features | — | Yes | Yes |
| Workspaces | 2 | Unlimited | Unlimited |
| Live broker integrations | — | 5 providers + simulator | 5 providers + simulator |
| Third-party providers | — | 12 across 6 sub-namespaces | 12 across 6 sub-namespaces |
| Autonomy tiers | Supervised | All 3 (Supervised / Guarded / Autonomous) | All 3 |
| Updates | Community | Priority | 3 years included |
Free forever. No credit card. No trial. The free version is a real AI workstation — not a demo.
Founder's Edition: One-time EUR 750. Limited to first 500 licenses. App works forever. When the 500th license sells, the price goes up — it will never drop.
We don't just say "we care about quality." We measure it. Every number below is re-checked before every release.
| Claim | Measurement | Value |
|---|---|---|
| Zero crashes from panics | .unwrap() that can actually panic in the production Rust tree |
0 |
| Zero technical debt hiding | #[allow(dead_code)] attributes anywhere in 572 files |
0 |
| Zero compiler warnings | cargo check across all 4 build modes (default, --features swarmforge, --tests, --tests --features swarmforge) |
0 |
| Zero build drift | Rust toolchain pinned via rust-toolchain.toml |
1.95.0 |
| Zero data leaks | Lines of code sending your data to us or anyone else | 0 |
| Every module stays healthy | Neural Trias (Memory + IPC + Health) covers every module | 313/313 |
| Every sub-module is emergent | Sub-modules declared via the Emergence Kernel | 82+ |
| Every gateway decision is logged | SHA-256 Merkle chain in the Autonomy gateway | every one |
| Every commit is auditable | Commits signed + reproducible (public release tags) | every one |
These aren't badges we bought. They are the output of cargo check,
grep, and git log on the public build — you can re-run any of them
yourself.
Most AI tools need expensive cloud APIs or 32 GB of RAM. ImpForge runs 4 specialized AI models on your machine — even an 8 GB laptop.
| Model | Role | What it does |
|---|---|---|
| qwen3-imp:8b | The Brain | Deep reasoning, strategic decisions, code generation. Selectively aligned (helps everywhere professional, refuses only violence/terrorism/fraud/exploitation). Has a "thinking mode" for hard problems. |
| Qwen2.5-Coder 1.5B | Fast Hands | Code completion in under 100ms. Fill-in-the-middle, syntax checking, refactoring. |
| nomic-embed-text | Memory | Turns your documents into searchable vectors. Powers the knowledge base and semantic search. |
| SmolLM2 1.7B | Classifier | Understands what you want in under 50ms. Routes your request to the right tool. |
Smart cascade routing: 60-70% of requests are handled by the tiny Classifier (instant). Only complex questions reach the Brain. This saves 3-5x compute.
Self-learning routing: The system tracks which model performs best for each type of task. After just 5 interactions, it automatically routes to the optimal model. It literally gets smarter the more you use it.
12-step quality pipeline: Every single AI response passes through privacy scanning, governance checks, quality validation, self-correction, PII scrubbing, cryptographic sealing, stigmergy trace, session context, evolution record, EU-compliance labelling, and module-memory feedback — automatically, invisibly, every time.
Board of Directors: For big decisions, the Brain simulates 10 executives (CEO, CTO, CFO, CMO, COO, CIO, CLO, CHRO, Analyst, Auditor) — each with a different perspective and temperature. They vote. You get a recommendation with dissenting opinions noted.
Works with every AI tool: 227 MCP tools organised into 13 namespace micro-servers (90% token savings vs monolithic). Claude Code, Cursor, JetBrains, Copilot, Windsurf — they all talk to ai_imp like a brother. Lazy schema loading: tool names first (~2K tokens), full schema on demand.
Selective Alignment: Your AI is helpful and safe. We don't ship a lobotomized model that refuses to write a port scanner. qwen3-imp helps with security research, creative writing, medical info, legal analysis — everything professional. But it firmly refuses violence, terrorism, fraud, and exploitation. No wholesale censorship. No wholesale uncensoring. Enterprise-grade balance.
# Linux
wget https://github.com/AiImpDevelopment/impforge/releases/latest/download/ImpForge.AppImage
chmod +x ImpForge.AppImage
./ImpForge.AppImage
# Windows — download ImpForge-setup.exe from Releases
# macOS — download ImpForge.dmg from Releases
ImpForge is made by AiImp Technology in Germany.
We believe AI should be:
Not like Grok. Like DeepSeek — open, democratic, accessible.
"The best AI is the one that respects you."
| Lines of code | 730,000+ |
| Automated tests | 6,742+ |
| Rust modules | 645 files, 322 modules |
| Sub-module files (focused micro-programs) | 300+ across 30+ directories |
| Free template gallery (MIT) | 78 industry × category templates — rolling out in 5 waves |
| Free skill library (MIT) | growing catalog of reusable AI agent skills |
template.json v1 specification |
published on GitHub — adopted by ImpForge, open to all |
| AI personalities | 472 |
| Quality rules | 158,000+ (81 AI-generated at benchmark 1.00) |
| Tauri commands | 2,975 |
| MCP tools (Claude Code / Cursor / JetBrains / Copilot / Windsurf) | 227 in 13 namespace servers |
| MCP token savings vs monolithic | 90% (lazy schema loading) |
| 4-Model AI Pipeline benchmark | 1.00/1.00 (all 6 dimensions perfect) |
| qwen3-imp:8b | Selectively aligned brain — PRODUCTION READY |
| Self-learning routing | Adapts from quality feedback automatically |
| Memory-informed routing | Module memory overrides classifier when quality > 0.85 |
| Neural Trias | Module Memory + Module IPC + Module Health (self-healing) |
| Emergence Kernel | One call wires every module's memory + IPC + health + MCP + insight bus |
| Lazy LSP Sidecars | 12 languages load only when you code them — Pyright's 1.7 GB only appears when you open a .py |
| Pipeline Modes | 7 (Streaming, Thinking, LongForm, Secure, Diverse, Code, Standard) |
| Widget Isolation | 10 built-in widgets, <5MB idle, ~1MB suspended |
| Project Builders | 12 categories (SaaS, Web, Mobile, Desktop, Firmware, CRM, E-commerce, Single-Page + sandbox + iterate + helpers + types) |
| Live broker integrations | 5 (Alpaca, IBKR, Binance, Kraken, Coinbase) + deterministic simulator |
| Third-party providers | 12 across 6 sub-namespaces (finance, content, cloud, payments, comms, social) |
| Autonomy tiers | 3 (Supervised / Guarded / Autonomous) with SHA-256 Merkle audit chain + fleet-wide kill switch |
| HyperChat-Next events | 11 AG-UI event variants (stream every LLM token + tool call + state change) |
| Autonomous micro-program modules | 400+ (each self-sufficient) |
| Module capabilities (each) | Privacy + Governance + Context + Stigmergy + Evolution + Memory |
| Anti-Palantir defense layers | 5 |
| DigiU privacy tiers | 5 (Fortress to Open) |
| Selective Alignment Engine | 8 safety categories REFUSED, 10 alignment categories HELPED |
| Neural synapses | 175 (40 in Synapse Fabric v2) |
| Specialized AI agents | 15 (Privacy, Compliance, Security, Governance, Evolution, ...) |
| Resource power modes | 5 (DeepSleep <80MB to Genesis 5GB VRAM) |
| Board of Directors | 10 executives via temperature variation + Hebbian trust voting |
| Governance rules | 23 (deny-by-default) |
| Emergent intelligence layers | 4 (Pattern, Anomaly, CrossModule, Summary) |
| ai_imp action types | 25 |
| A2UI components | 20+ |
| Design tokens | 400+ |
| Research papers cited | 608+ |
| EU AI Act features classified | 17 (all compliant) |
| Build errors | 0 |
| Build warnings | 0 |
Production .unwrap() |
0 (Crown-Jewel grade) |
#[allow(dead_code)] |
0 anywhere |
| Rust toolchain | Pinned 1.95.0 (reproducible builds) |
| Data collected from you | 0 |
ImpForge — DigiU
Your desktop. Your AI. Your rules.
Download •
Feedback •
Build from Source