≈AXF
Status: v0.1 Draft — Open for comment

AXF — an open wire format for AI agents.

60-80% fewer tokens than JSON. Human-debuggable. Stream-native.

Inspired by 40 years of battle-tested EDI density. Built for LLM agents.

Read the spec → See examples Downloads

Token Efficient

No structural punctuation tax. Field positions defined by schema, not repeated on the wire.

Human Readable

ASCII-only, line-oriented, debuggable with cat. If you can read EDI, you can read AXF.

Stream Native

Segment-terminated. Parsers emit events as they go. No need to buffer the whole message.

Schema Driven

Envelope names a schema ID. Receivers already know the layout. Zero key overhead.

Why?

Same tool call. Left: JSON. Right: AXF. Numbers use the cl100k_base tokenizer (GPT-4 family); other tokenizers differ but the direction holds.

tool-call.json~135 tokens · 414 bytes

{
  "formatica_version": "1.0",
  "schema": "tc:weather.v1",
  "request_id": "8a3f",
  "header": {
    "role": "assistant",
    "timestamp": 1745515200,
    "conversation_id": "c_9b2e"
  },
  "tool_call": {
    "name": "get_weather",
    "arguments": {
      "location": "Austin, TX",
      "units": "imperial",
      "when": "now"
    }
  },
  "metadata": {
    "priority": "normal",
    "timeout_ms": 5000
  }
}

tool-call.axf~58 tokens · 164 bytes · ~57% smaller

FX*1.0*tc:weather.v1*req:8a3f~
H*role:assistant*ts:1745515200*conv:c_9b2e~
TC*get_weather*loc:Austin,TX*units:imperial*when:now~
M*priority:normal*timeout:5000~
E~

Why the savings? JSON spends a lot of tokens on {, }, ", ,, and whitespace. AXF uses single ASCII delimiters that tokenizers treat as cheap separators, plus schema-by-reference so keys don't ride along on every message.

Get involved

AXF is an open standard. The spec is a draft — feedback, test vectors, and reference implementations are welcome.