Welcome to the Agentic Age. No question is bad. No topic is off limits. We are here for you — to help every engineer ship faster, think bigger, and build with confidence.
Claude Opus 5 is now default, 1M context
Background sessions now commit & push work
Emoji shortcode autocomplete in prompts
Fixed --mcp-config servers not connecting before first turn in print mode
Fixed memory leak in truncated MCP tool outputs
Subagents nest to depth 3 by default
Fixed quadratic slowdown in long-session message normalization
Fixed auto-compact never triggering for Opus 4.8 on Bedrock
Token-Oriented Object Notation — a compact, structured translation layer for passing data to LLMs without wasting tokens.
TOON stands for Token-Oriented Object Notation — a lightweight format purpose-built for LLM workflows where every token counts.
TOON represents the same structured data as JSON — objects, arrays, fields, and values — but in a far more compact form.
Designed specifically for LLM prompts where token usage drives both cost and context window consumption.
Despite its compactness, TOON stays readable. Humans can inspect it, models can parse it — no custom decoder needed.
JSON repeats field names for every single object
In LLM prompts, repetition = wasted tokens = higher cost + less room for content
In a list of 100 users, the keys "id", "name", and "role" appear 100 times each — adding hundreds of tokens that carry zero new information.
Extra tokens increase API cost directly. On large datasets — catalogs, forecasts, transaction records — JSON overhead compounds fast.
Every token spent on syntax is a token stolen from actual content. Verbose payloads leave less room for instructions, examples, and reasoning.
TOON is compact and structured — a rare combination that makes it uniquely suited to model consumption.
Eliminating repeated keys directly reduces token count. On large uniform datasets the savings are substantial — less spend per call, more headroom for actual content.
Uniform row-based data is easier for a model to scan and reason over than deeply nested JSON trees. A consistent pattern reduces ambiguity.
The declared record count — [2] in users[2] — lets a model (or your code) detect missing or truncated data before processing begins.
The field list at the top acts as a schema header. Downstream code and the model both know exactly what columns to expect — no guessing from context.
TOON works best with uniform arrays of objects — repeated records that share the same fields.
Passing user records, roles, or permission sets to a model for analysis, filtering, or summarization.
SKU lists, pricing tables, inventory snapshots — any catalog with consistent fields across many items.
Time-series data, KPI snapshots, and trend tables that share a consistent schema across every row.
Financial transactions, event logs, audit trails — high-volume records where JSON overhead compounds quickly.
Ranked result sets returned from a search API — consistent structure, many records, ideal for TOON encoding.
If your data is an array of objects where every object has the same fields, TOON is almost certainly a better fit than JSON.
TOON is a specialized tool — not a universal JSON replacement. Reach for it only when it adds clear value.
TOON's row-based format assumes flat, uniform records. Deeply nested objects or records with variable structure don't map cleanly to TOON's schema-then-rows model.
If objects in your array don't share the same keys, TOON's single-schema header breaks down. Use JSON when field sets vary per record.
If your data is a simple flat table with no nesting and no metadata, plain CSV is even more compact than TOON and universally understood.
For small payloads (a single object, a handful of records) the overhead of switching formats outweighs the token savings. Don't over-engineer small data.
Keep using JSON. TOON is a translation layer — your application logic stays the same.
Convert to TOON before passing structured data to a model. Fewer tokens, same meaning.
Repeated records with consistent fields — user lists, catalogs, forecasts, transactions.