Aether

A deterministic orchestration language for AI agents.

---

LLM agents fail silently, run code with no safety model, and produce pipelines that are impossible to audit. Aether is an intermediate representation (IR) that gives AI-generated pipelines verifiable intent, typed outputs, and compile-time safety gates — without replacing the guest languages (Python, JS, shell) your nodes already use.

"Intent is Compilation."

How it works

You write (or generate) a .ae program describing a pipeline as a directed acyclic graph of typed nodes. Each node declares what it does (_intent), what safety level it requires (_safety), what it reads (::IN), what it writes (::OUT), and what must be true after it runs (::VALIDATE). The Aether kernel executes the DAG, enforces safety gates, validates outputs, and — if a node fails — optionally calls an LLM to repair the code and retry.

§ROOT 0xFF_MAIN {

  §ACT 0x1A {
    ::META { _intent: "fetch active users", _safety: "read_only" }
    ::EXEC<PYTHON> {
      import urllib.request, json
      return json.loads(urllib.request.urlopen("https://api.example.com/users").read())
    }
    ::OUT { $0xUSERS: Type<JSON> }
  }

  §ACT 0x2B {
    ::META { _intent: "filter to adults only", _safety: "pure" }
    ::IN   { $0xSRC: Ref($0xUSERS) }
    ::EXEC<PYTHON> {
      return [u for u in $0xSRC if u["age"] >= 18]
    }
    ::OUT      { $0xADULTS: Type<JSON> }
    ::VALIDATE { ASSERT len($0xADULTS) >= 0 OR HALT }
  }

}

Referencing ledger state in ::EXEC

Any $0xADDR token in guest code resolves from the state ledger at execution time. You can either:

• Declare explicitly with ::IN (recommended) — gets you typed inputs, ghost aliases ($0xSRC [users]), and a documented dependency edge in the audit log.
• Reference directly without ::IN — the kernel scans the code for any $0x* addresses not already bound and pulls them from the ledger. Handy for one-liners that read a single ::CTX value. You lose the typed-binding guarantees but the pipeline still runs.

Both forms produce the same ledger reads; ::IN just makes the contract explicit.

Or use Aether-Short (.as) for ~60% fewer lines:

@pipeline 0xFF_MAIN

  $0xUSERS:  JSON = @std.io.net_get() {
    import urllib.request, json
    return json.loads(urllib.request.urlopen("https://api.example.com/users").read())
  }

  $0xADULTS: JSON = @std.proc.list.filter($0xUSERS) {
    return [u for u in $0xUSERS if u["age"] >= 18]
  } | ASSERT len($0xADULTS) >= 0 OR HALT

@end

Why not just write Python?

	Raw Python / LangGraph	Aether
Safety model	None	L0–L4 compile-time gates
Typed outputs	Runtime duck typing	Declared + validated at write time
Audit trail	Manual logging	Automatic output.ae.json with full node traces
Self-healing	Manual try/except	ASSERT ... OR RETRY(3) — LLM repairs the node
LLM generation cost	715–1,225 tokens (LangGraph/AutoGen)	415 tokens (3-node baseline)
Runtime LLM calls	0 (LangGraph) — 6,200 (AutoGen)	0 for deterministic pipelines

See docs/benchmark.md for the full token-cost comparison.

Features

Safety — the declaration is checked, not trusted

• 5-tier safety model — L0 (pure math) → L4 (system root). Nodes above your threshold are blocked before execution.
• Effect inference — the kernel reads each node's guest code and derives the level it actually needs. A mismatch is reported on every run; --verify-effects blocks it. A node tagged pure that imports requests does not get to run.
• Secret taint propagation — values resolved from ENV: are redacted, and so is anything computed from them, in the ledger, the trace and the cache alike. std.sec.hash.sha2 and std.sec.mask.pii are the sanctioned ways to release a derivative.

Verifiability — a run you can check afterwards

• Execution root hash — every node is content-addressed over its code, resolved inputs and declared contract; the per-node commitments fold into one root_fingerprint. Two runs sharing that hash provably ran the same code on the same inputs and got the same answers.
• Plan validation — aether check <file> (or --dry-run) proves every ::IN reference resolves, no address has two writers, and the graph is acyclic. Exits non-zero, so it drops straight into a save hook or CI.
• Typed state ledger — outputs are written to an address space ($0xADDR) with type validation on every write.

Execution

• Parallel DAG execution — independent nodes run concurrently, bounded by --max-parallel; dependencies resolved with Kahn's topological sort.
• §MAP fan-out — §MAP 0x2B OVER $0xITEMS AS $0xITEM { … } runs one body per element of a runtime list. A map is one node in the graph, so plan checking and the root hash keep working even though N is unknown until the run.
• §FAIL handlers — a block that runs after the graph, and only if something failed, with the failure summary bound to $0xFAILURES. Write a partial report, notify, clean up. ::ON { 0x2B } scopes a handler to specific nodes, and scopes what it sees, so reacting differently to different failures does not mean branching inside the guest.
• Memoization — --cache reuses results for nodes whose code and inputs are unchanged. Eligible only when pure/read-only, statically transparent, and untainted; a cached run still agrees with an uncached one on the root hash.
• Per-node timeouts — --timeout bounds any single node, and the guest process is killed rather than abandoned.

Authoring

• ASL registry — 38 canonical intents (std.io.*, std.proc.*, std.ml.*, std.sec.*, …) with safety and language defaults. Unknown std.* intents are hard errors; custom namespaces (myorg.etl.*) are accepted.
• Self-healing RETRY — ASSERT expr OR RETRY(3) sends the failing code to a registered LLM handler for repair. It remembers every rejected attempt, refuses assertions no rewrite could satisfy, and rejects candidates that do not parse before spending a call.
• English Toggle — aether gen "description" turns plain English into a .ae program.
• MCP server — aether-mcp exposes validate/execute/audit/inspect as tools for Claude Code and other MCP-compatible clients. aether_validate runs the full plan check, not just a parse, so an agent checking its own generated pipeline gets the same answer aether check would give.
• REST API — aether-api runs on port 3737 for LangChain, AutoGen, n8n, or any HTTP client. GET /grammar returns a machine-readable summary of the language for LLM consumption.
• Aether Lens — a standalone DAG visualizer (lens/index.html) that renders any output.ae.json execution log.

On isolation. Guest code runs as an ordinary host subprocess. Everything above is analysis — it catches accidents and dishonest safety tags, which is the common case, but it does not contain a determined program. Real containment needs the WASM runtime tracked in ROADMAP.md. Run untrusted pipelines under an OS-level sandbox of your own until then.

Aether Lens

Every execution writes an output.ae.json audit log. Open lens/index.html in any browser to visualize it — no build step, no npm, no server required.

Each node card shows its intent, safety level, execution time, and status — color-coded at a glance:

Color	Meaning
Green	Completed successfully
Amber (HEALED)	Failed, repaired by LLM, re-ran successfully
Red (BLOCKED)	Skipped — upstream dependency failed or safety gate rejected
Orange	Failed with no recovery

Click any node to open the detail panel: full JSON output, validation results, and the heal log if self-healing was attempted. The ledger bar at the bottom shows every live state address and its current value.

# Serve locally to auto-load output.ae.json
cd lens && python -m http.server 8080
# → open http://localhost:8080

# Or just open the file directly and drag-and-drop any output.ae.json
open lens/index.html

Installation

# Option 1: pip (recommended — no Rust required)
pip install aether-kernel

# Option 2: Pre-built binary
curl -fsSL https://raw.githubusercontent.com/baiers/aether/main/install.sh | bash

# Option 3: Build from source
cargo build --release

Quick Start

# Run a pipeline
aether examples/demo.ae

# Expand Aether-Short to inspect generated .ae
aether examples/pipeline.as --expand-only

# Run with a higher safety threshold (allow network calls)
aether examples/demo_showcase.ae --safety l3

# Self-healing (requires ANTHROPIC_API_KEY)
ANTHROPIC_API_KEY=sk-... aether examples/self_heal_demo.ae

# Generate a .ae program from plain English
ANTHROPIC_API_KEY=sk-... aether gen "fetch the top 10 HN stories and summarize each one"

# Start the REST API
aether-api  # → http://localhost:3737

# Start the MCP server (for Claude Code integration)
aether-mcp

Claude Code / MCP Integration

Add to your MCP config (~/.claude/mcp_config.json):

{
  "mcpServers": {
    "aether-kernel": {
      "command": "aether-mcp",
      "args": []
    }
  }
}

Claude can then call these directly:

Tool	Does
aether_validate	Parse and plan-check without running: every ::IN Ref() resolves, no address has two writers, the graph is acyclic, every §FAIL ::ON target exists
aether_execute	Run a program. Takes safety_level, verify_effects, strict_registry, use_registry, timeout_seconds, max_parallel
aether_audit	Send an execution log to the registered LLM handler for a natural-language report
aether_inspect	Retrieve a previous execution log by its id

Set verify_effects: true when running code you did not write — it infers each node's real effect level from its guest body and blocks any node needing more than its _safety tag declares.

Project Structure

src/          Rust kernel (parser, executor, ASL registry, self-healing, MCP, REST API)
asl/          ASL registry — 38 canonical intents (JSON)
examples/     Runnable .ae and .as programs
spec/         Formal EBNF grammar and type system docs
docs/         Whitepaper, benchmark paper, kernel manual
lens/         Aether Lens DAG visualizer (standalone HTML, no build step)
sdk/          LLM system prompt, MCP config, audit prompt
benchmark/    LangGraph / AutoGen equivalents + token counting scripts
python/       pip-installable wrapper (aether-kernel)

What's Included (Community — Free, Apache 2.0)

Feature
Parser, AST, topological executor	✓
5-tier safety model (L0–L4)	✓
ASL registry (38 canonical intents)	✓
Aether-Short (.as) notation	✓
Self-healing RETRY (your ANTHROPIC_API_KEY)	✓
English Toggle aether gen (your ANTHROPIC_API_KEY)	✓
MCP server for Claude Code	✓
REST API	✓
Aether Lens DAG visualizer	✓

Aether Pro (coming soon) — hosted execution, managed LLM calls (no API key needed), extended ASL (200+ intents), persistent history. Join the waitlist →

License

Licensed under the Apache License 2.0.

See docs/open-core.md for the Community vs Pro vs Enterprise breakdown.