ANAL

ANAL (Agent Notation, Actually Lean) is a compact, token-lean serialization format for data flowing between agents and models. It needs no model retraining: it wins by matching the tokenizer rather than fighting it.

Why

LLMs bill by tokens, not characters. JSON repeats field names, quotes, and brackets for every value. ANAL declares field names once in a header and then sends only values.

Install

go get github.com/SomeoneUnlicensed/anal

Quick start

package main

import (
	"fmt"

	"github.com/SomeoneUnlicensed/anal"
)

func main() {
	rows := []map[string]any{
		{"id": 1, "sku": "P123", "qty": 2, "active": true},
		{"id": 2, "sku": "P124", "qty": 1, "active": false},
	}
	data, _ := anal.Marshal(rows)
	fmt.Print(string(data))
}

Output:

@table: {active,id,qty,sku}
T|1|2|P123
F|2|1|P124

Format

@table: {id,sku,qty,price,active}
1|P123|2|100|T
2|P124|1|150|F

@obj: {event,order_id,status,total}
order.created|8123|paid|199.9

@list:
code
coffee

• @table — a list of flat objects, one |-separated row each.
• @obj — a single flat object.
• @list — a flat list of scalars, one value per line.

Values: T / F for booleans, ~ for null, numbers as-is, strings as-is unless ambiguous or containing special characters — then backtick-quoted.

Escapes: \\ \` \| \n \t \r.

API

Marshal accepts the same kinds of values as encoding/json — structs, maps, slices, and scalars — and honors json struct tags. Unmarshal does the reverse.

Nested objects and arrays are not supported in v1 and return an error. Use JSON for deeply nested data.

Integrate

Install the CLI:

go install github.com/SomeoneUnlicensed/anal/cmd/anal@latest

Install the ANAL skill so coding agents know to compress large structured data:

anal install          # Codex + Claude Code
anal install codex    # Codex only
anal install claude   # Claude Code only

echo '[{"id":1,"sku":"P123","qty":2}]' | anal enc

Python helpers live in tools/:

• analctl.py — convert JSON <-> ANAL and diff to measure token savings.
• agent_shim.py — wrap_tool_result() to drop compact tool results into context, and decode() to read them back.

python tools/analctl.py diff < payload.json

Benchmark

Token counts for the same data, measured with tiktoken (cl100k_base):

encoding	chars	tokens
JSON	133	61
TOON	81	38
ANAL	63	35

CPU benchmarks:

go test -bench=. -benchmem

License

MIT — see LICENSE.