Zixir

<p align="center"><img src="assets/zixir-icon.png" alt="Zixir" width="128" /></p>

Repository: github.com/Zixir-lang/Zixir · Created and maintained by Leo Louvar.

Zixir is an immutable AI automation language that bakes workflow orchestration, resource limits, and observability into one runtime—no Airflow + Redis + Prometheus glue. Small, expression-oriented, on a three-tier runtime: Elixir (orchestrator), Zig (engine), Python (specialist).

Zixir is its own language (own grammar and semantics), implemented with Elixir, compiling to Zig, and calling into Python. You write .zixir source; it is not Elixir or Zig syntax. Zixir is immutable by design: variables cannot be reassigned, so code is easier to reason about and less prone to bugs than in mutable-by-default languages.

Who it's for: Developers and teams building AI automation, agentic workflows, and ML pipelines who prefer a single, expression-oriented language and runtime over managing Airflow, K8s, Redis, and custom YAML. Best fit for engineers who like Elixir/FP, want pattern matching and type inference, and need built-in fault tolerance and observability without extra infra.

Why Zixir?

Problem: Production AI pipelines usually need Airflow, K8s, Redis, Prometheus, custom fault-tolerance code, and YAML.

Solution: One language. Orchestration, checkpointing, resource limits, circuit breakers, and built-in caching (ETS + disk, no Redis) are in the runtime.

| Strength | Trade-off | |----------|-----------| | Immutable by default (single-assignment; no in-place mutation) | — | | Built-in caching (ETS + disk) | — | | Pattern matching (native; unique among workflow tools) | — | | Interactive REPL | — | | Fault tolerance (supervision, circuit breakers, retries) | — | | Less infra, faster dev, type safety | Learning Elixir/Zixir; smaller ecosystem; newer project |

Why a three-tier runtime?

Each tier does one job well; together: orchestration, speed, and ecosystem.

| Tier | Role | Strength | |------|------|----------| | Elixir | Orchestrator | Concurrency, fault tolerance, supervision (“let it crash”), OTP. Coordinates tasks and keeps the system up. | | Zig | Engine | Predictable performance, no GC, small binaries. Hot paths: parsing, math, core ops (NIFs). | | Python | Specialist | ML (PyTorch, TensorFlow), data (pandas, numpy), APIs. Use existing libraries without rewriting. |

Elixir orchestrates and restarts failed workers; Zig runs hot-path code; Python handles ML and data.

How it works

Zixir has its own grammar (let, expressions, array/map indexing arr[i] and map["key"], engine.op(args), python "module" "function" (args), literals, binary ops, 25+ built-in functions). Source is parsed in Elixir into a Zixir AST, then either:

• Interpreted — Zixir.eval(source) evaluates the AST in Elixir; engine.* calls run in Zig NIFs, python calls go to Python via a port.
• Compiled — Zixir.Compiler.compile(source) type-checks, optimizes, and emits Zig; the Zig is compiled to a native binary or run JIT.

Your code stays Zixir-only; the runtime is Elixir + Zig (engine/codegen) + Python (libraries).

Zixir vs. alternatives (honest assessment)

| Feature | Zixir | Airflow | Kubeflow | Prefect | |---------|-------|---------|----------|---------| | External infrastructure | Elixir runtime only* | Redis + DB | Kubernetes | Minimal | | Setup time | ~20 min | ~2 hours | ~2 days | ~1 hour | | Lines of code (typical ML pipeline) | 50–150 | 200–500 | 500–1000 | 150–400 | | Workflow orchestration | ✅ Built-in | ✅ | ✅ | ✅ | | Checkpointing | ✅ Any type | ⚠️ JSON only (XCom) | ✅ | ✅ | | Resource limits | ✅ Code-level primitives | ✅ Config (e.g. execution_timeout) | ✅ YAML | ✅ Decorators | | Fault tolerance | ✅ Supervision + circuit breakers | ⚠️ Basic | ⚠️ K8s | ⚠️ Basic | | Observability | ✅ Built-in | ⚠️ Manual | ⚠️ Complex | ⚠️ UI | | Pattern matching | ✅ Native | ❌ | ❌ | ❌ | | Interactive REPL | ✅ | ❌ | ❌ | ❌ | | Type inference | ✅ | ❌ | ❌ | ❌ | | Immutable (no reassignment; safer, easier to reason about) | ✅ | ❌ | ❌ | ❌ | | Native performance | ✅ Zig NIFs | ❌ | ❌ | ❌ | | LSP Support | ✅ mix zixir.lsp | ❌ | ❌ | ❌ |

* Elixir + Zig (build-time) required; no Redis, K8s, or separate DB for workflows.

Bottom line: One language replaces the usual Airflow + Redis + Prometheus stack. Requires Elixir and Zig; once set up, you get orchestration, limits, and observability without 3–5 external services.

Layout (three-tier flow)

flowchart TB
  Source[Zixir source / eval]

  subgraph T1["Tier 1: Elixir - Orchestrator"]
    Intent[Intent / routing]
    Memory[Memory / state]
    Supervisor[Supervision]
  end

  subgraph T2["Tier 2: Zig - Engine"]
    NIF[NIFs]
    Math[Math, parsing, core ops]
  end

  subgraph T3["Tier 3: Python - Specialist"]
    Port[Port bridge]
    Libs[ML, data, scripts]
  end

  Source --> Intent
  Intent --> Memory
  Intent --> Supervisor
  Intent -->|hot path| NIF
  Intent -->|library calls| Port
  NIF --> Math
  Port --> Libs

📚 Zixir Language complete guide

Zixir Language complete guide — Learn the language from zero to real projects. This is what the complete guide shows:

What's inside

| Part | Content | |------|---------| | Part 1: Getting Started (1–5) | Installation (all platforms), your first program, REPL, Hello world with explanations | | Part 2: Language Fundamentals (6–12) | Variables and all data types (Int, Float, Bool, String, Arrays, Maps), map indexing, type conversions, built-in functions, exercises | | Part 3: Control Flow (13–17) | If/else, while and for loops, pattern matching, decision-making patterns | | Part 4: Functions (18–22) | Definition and calling, parameters vs arguments, lambdas, recursion, scope and best practices | | Part 5: Power Features (23–28) | 25+ built-in functions, pipe operator (\|>), engine operations (22 Zig NIFs), Python integration, performance | | Part 6: Real-World Projects (29–35) | Data pipeline, AI text analysis, LLM integration, workflow automation | | Part 7: Mastery (36–40) | Best practices, performance tips, debugging guide, patterns and anti-patterns | | Appendices | Grammar reference, quick reference card, common patterns, engine operations table |

Key features: Progressive learning, 20+ complete examples, 4 real projects, exercises with solutions, visual aids, error-handling focus, AI/automation emphasis.

Guide stats: 40+ pages · 20+ code examples · 4 projects · 6 exercises per chapter · 25+ built-in functions · 22 engine ops · pipe operator, modulo, map indexing, type conversions · full grammar reference.

PDF: A PDF copy is available at docs/Zixir Language complete guide.pdf (same path as the guide, .pdf for download). Website: zixir-lang.github.io/Zixir/Zixir%20Language%20complete%20guide.pdf. To rebuild locally: Node — npx md-to-pdf "docs/Zixir Language complete guide.md" (output is already named with spaces); or pandoc — ./scripts/build-guide-pdf.sh (Unix) or .\scripts\build-guide-pdf.ps1 (Windows; requires pandoc and LaTeX). To enable automatic PDF build on push: create .github/workflows/build-guide-pdf.yml from scripts/build-guide-pdf-workflow.yml (skip the first 3 comment lines).

Requirements

| Requirement | Notes | |-------------|--------| | Elixir 1.14+ / OTP 25+ | Runtime | | Zig 0.15+ | Build-time only; run mix zig.get after mix deps.get (Zigler) | | ~100 MB disk, file system | For persistence, checkpoints, cache | | Python 3.8+ (optional) | For ML/specialist calls; set config :zixir, :python_path if not on PATH |

Platforms: Windows, macOS, Linux. Zigler compiles NIFs at mix compile; Python script: priv/python/port_bridge.py.

Optional: MLIR (Phase 4) — For extra optimizations (vectorization, CSE, constant folding) you can add the Beaver dependency on Unix: {:beaver, "~> 0.4"} in mix.exs. Without it, the compiler still runs AST-level optimizations. See docs/MLIR_AND_PYTHON.md. Windows: Beaver/Kinda not supported; use default (no Beaver).

New to Zixir? For step-by-step install of Elixir, Zig, and Python per OS, see SETUP_GUIDE.md.

Quick start

First-time install

git clone https://github.com/Zixir-lang/Zixir.git
cd Zixir
git checkout v7.1.0
mix deps.get
mix zig.get   # after deps.get, for Zigler
mix compile
mix phx.server   # start web UI

Then open http://localhost:4000 in your browser.

Clean build (troubleshoot or reinstall)

Use this when you need a clean build, are troubleshooting, or reinstalling:

# 1. Clone the repository
git clone https://github.com/Zixir-lang/Zixir.git

# 2. Navigate into the project directory (CRITICAL!)
cd Zixir

# 3. Reset any local changes and fetch tags
git checkout -- .
git fetch origin --tags
git checkout v7.1.0

# 4. Full clean
mix clean
mix deps.clean --all

# 5. Get dependencies FIRST
mix deps.get

# 6. Install Zig
mix zig.get

# 7. Compile
mix compile

# 8. Start server
mix phx.server   # start web UI

Then open http://localhost:4000 in your browser.

For Python specialist: ensure Python is on PATH or set in config; recommend a virtualenv. For VectorDB (nine backends: memory, chroma, pinecone, weaviate, qdrant, milvus, pgvector, redis, azure), see docs/VECTORDB_BACKENDS.md for pip install and setup.

Usage

Zixir language (source)

Run Zixir source with eval/1 or run/1:

Zixir.eval("engine.list_sum([1.0, 2.0