aiCompiler — The first programming language where the LLM is the CPU
AI aicompiler.devTry it free → Not a framework. Not an agent wrapper. A new computing paradigm.
The first virtual machine
powered by an LLM.
You write .aic — plain Markdown, no code.
aiCompiler resolves all ambiguity at compile time.
**aiVM executes intent opcodes using an LLM — just as the JVM executes bytecode using a CPU.**
LangChain / LangGraph → orchestrate LLM calls in Python code
aiCompiler → write intent in Markdown. aiVM is the runtime. LLM is its engine.
.aic→aicompiler compile→.aix→aivm run→✓ output
| Era | Language | Compiler | VM | Engine | | --- | --- | --- | --- | --- | | 1950s | Assembly | Assembler | — | Silicon | | 1990s | Java | javac | JVM | CPU | | 2026 | .aic | aiCompiler | aiVM | LLM |
🖥️
1950s — 2025
Von Neumann Computer
Silicon CPU executes machine code instructions sequentially
CPU
Executes machine code
RAM
Program state + stack
Cache
Fast recent-access storage
Registers
Working variables
System Bus
Connects all components
Hard Disk
Persistent storage
ROM / BIOS
Read-only firmware
🤖
2026 →
aiVM Computer
LLM executes intent opcodes — each step is a direct model inference
LLM Engine
Executes intent opcodes
Context Window
Working memory per turn
.aix Lock Files
Cache — same spec = instant
Step State
$step_1.output, $input.x
Tool Adapters
DB, email, APIs, HTTP
KV + VectorDB
Persistent memory + RAG
System Prompt
Read-only, loads at boot
CPU≡LLM Engine
RAM≡Context Window
Cache≡.aix Lock Files
Registers≡Step State
System Bus≡Tool Adapters
Hard Disk≡KV + VectorDB
ROM / BIOS≡System Prompt
Try it right now
Write a workflow. Watch aiCompiler compile it to semantic opcodes. Watch aiVM execute each step — the LLM reading opcodes like a CPU. No account. No install. No API key.
Examples:
Real AI — Llama 3.3 70B
Click Run to execute
the workflow step by step
.aic→aicompiler compile→.aix→aivm run→step log
How it works
Three concepts. That's all you need to know.
01
✍️
source
Write a .aic file
Write your program in plain English — workflows, steps, tools. No semicolons. No type annotations. No boilerplate.
Workflow: Greeter
Steps
1. SayHello
Say hello to {name} creatively.
02
⚡
compile
Compile to .aix
The AI compiler parses your source file and produces bytecode — structured JSON opcodes with zero ambiguity.
aicompiler compile app.aic → app.aix ✓
03
🚀
run
Run with aivm
The AI VM reads each opcode and executes it — calling real tools, resolving data references, returning structured output.
aivm run app.aix → Hello, World! ✓
Why developers care
The framework era is ending. This is what comes next.
⚙️
LLM is the CPU
The LLM doesn't assist your code. It IS the runtime. Each opcode is a direct instruction — no Python glue, no chain configuration.
🔀
Swap any model
Claude, Nova Micro, Gemini, Llama — change one flag. All routed through Cloudflare AI Gateway. No rewrite needed.
🔒
Deterministic builds
.aix lock files ensure the same source always compiles identically. Reproducible like javac. Version-controlled like code.
⚡
Enrich between steps
Two-step pipeline like Java. Between compile and run: inject tool bindings, retry policies, security rules, provider selection per step.
🐛
Debuggable by design
Every step produces a structured log: name, inputs, outputs, duration, error handling. Not a blob of LLM output.
🛠️
Open source, MIT
Fork it. Build your own AI language on top of it. The .aix opcode format is an open spec.
A new kind of VM.
Try it in 30 seconds.
No install. No account. No API key. Just open the playground and run.