An experimental project from The Weekend Projects: an AI-native programming language that treats the LLM as the CPU. You write what you want to do in plain Markdown (.aic), the aicompiler toolchain resolves the ambiguity into semantic opcodes expressed as JSON (.aix), and the aiVM runtime executes them one step at a time with an LLM as its engine. It is the JVM-running-bytecode-on-a-CPU picture, with the CPU swapped out for an LLM. The key distinction is that the LLM is not called as a “tool” from inside conventional code — it sits at the execution layer itself, which is why the project bills itself as a new computing paradigm rather than a framework or an agent wrapper. It is released as MIT-licensed open source, and the official site offers a live demo that needs no account, no API key, and no installation.
Key Features
.aicsource written in plain English: No semicolons, no type annotations. You describe a workflow’s inputs, the tools it uses, and its steps in Markdown-style natural language, and that is the program- Compilation to
.aixpins down the ambiguity: Instructions that would be interpreted differently on every run in raw natural language are converted at compile time into a zero-ambiguity sequence of opcodes (JSON). The output is human-readable, so it can be committed to Git and reviewed as a diff like any other artifact - aiVM executes step by step: The aiVM runtime processes the
.aixin order, calling external tools along the way and managing state between steps. The structure makes it possible to trace what happened at which stage - Swappable models: The default is Llama 3.3 70B on Cloudflare Workers AI. Anthropic’s Claude and Amazon Bedrock’s Nova Micro are also supported, so the same
.aiccan be run on a different model - Runs on Cloudflare Workers: You clone the repository and deploy it to Cloudflare Workers. The web interface is built with Astro and React
- MIT-licensed open source: The code is public on GitHub, free to read, modify, and run in your own environment
Pricing
| Plan | Price | What you get |
|---|---|---|
| Live demo on the official site | $0 | No account, no API key, no installation. Try the sandbox samples on the spot |
| Self-hosting (MIT license) | $0 | Clone from GitHub and deploy to Cloudflare Workers |
| LLM usage | Provider pay-as-you-go | The default Cloudflare Workers AI (Llama 3.3 70B) has a free tier. Using Claude or Amazon Bedrock incurs each provider’s own charges |
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- You can build a workflow by writing what you want in English, without learning a programming language’s syntax
- The compiled
.aixis human-readable JSON, so you can review how the AI interpreted your intent — more transparent than throwing a whole prompt at a model - Because
.aixcan be version-controlled, it helps contain the problem of instructions being reinterpreted differently on every run - The same source carries over when you swap models, so you are less locked into model-specific phrasing
- Completely free under the MIT license, and you do not even need an account to try it
⚠️ Cons
- This is an experimental project at v0.1; external tool integration is still under development, and there is no production track record yet
- It was published recently, and both the GitHub star count and the community are small. There are almost no articles or case studies to fall back on when something goes wrong
- Because the execution engine is an LLM, tightening ambiguity at compile time does not make the output fully deterministic
- The setup assumes deployment to Cloudflare Workers, so running it elsewhere requires modification
- Both the documentation and
.aicauthoring assume English; there is essentially no Japanese-language material
Comparison with Similar Services
| Criteria | aiCompiler | DSPy | BAML | LangChain |
|---|---|---|---|---|
| Positioning | AI-native language + virtual machine | Python framework for writing LLM programs declaratively | DSL for defining LLM functions with typed client generation | General-purpose framework for building LLM apps |
| How you write | Plain Markdown (.aic) | Python code | Dedicated DSL (.baml) | Python / TypeScript code |
| Intermediate representation | .aix (JSON opcode sequence) | None | Generated typed code | None |
| Maturity | Experimental, v0.1 | Research-born OSS with production use | Production use | The widest production adoption |
| Price | Free (MIT-licensed OSS) | Free (OSS) | Free (OSS) | Free (OSS, with paid surrounding services) |
DSPy, BAML, and LangChain all follow a “call the LLM from inside an existing programming language” design. aiCompiler stands apart by replacing the language itself with natural text and placing the LLM at the execution layer. If you are choosing on practicality, a mature existing framework is the sensible pick — but aiCompiler is the more interesting one as a design idea.
Who Is It For
- Developers curious about treating the LLM as a computing substrate rather than a tool to call
- Anyone frustrated by the poor reproducibility of piping prompts straight into execution, who wants to try an approach with an intermediate representation
- People who want to experiment with running AI workflows on Cloudflare Workers
- Anyone who likes following a new programming paradigm from its earliest stage, reading the code as it evolves
Summary
aiCompiler is an experimental project trying to answer, with working code, the question of what a language, a compiler, and a VM look like once you cast the LLM as the CPU. The idea of compiling natural-language source into an intermediate representation you can version-control also speaks to the very practical problem of prompt reproducibility. But at v0.1, with little track record and little written about it, the right way to approach it is not as a tool for production work — it is as something to experience through the live demo on the official site, to see what the idea feels like.