Harness Context-Budget Planner
See how many context-window tokens your CLAUDE.md/AGENTS.md, skills, and tool schemas consume before a user types anything — split into always-on vs on-demand, with trim suggestions.
| Model | Context Window (tokens) |
|---|
See how many context-window tokens your CLAUDE.md/AGENTS.md, skills, and tool schemas consume before a user types anything — split into always-on vs on-demand, with trim suggestions.
| Model | Context Window (tokens) |
|---|
No. Everything you paste is counted and analyzed in your browser with JavaScript. Nothing is sent to a server, logged or stored, and closing the tab discards it.
They are estimates from a characters-per-token ratio you can change. The default of 2.5 is the conservative figure Anthropic documents for current Claude models (about 2.5 million characters per 1 million tokens); 4 is the usual English-prose rule of thumb for GPT and Gemini tokenizers. Code, JSON and non-English text usually need more tokens per character. Use the numbers to compare parts of a harness and spot bloat. For exact counts use the provider's counter: Anthropic's count_tokens endpoint, OpenAI's tiktoken, or Gemini's countTokens.
Always-on tokens are in the context window on every turn: your instructions file (CLAUDE.md or AGENTS.md), tool and function schemas, and the short name and description of each skill. On-demand tokens, mainly a skill's full body, load only when the skill is used. Always-on size is what you pay for before the user types anything and what competes with the task for the model's attention.
As small as it can be while still carrying what the agent cannot discover from the repository itself: build and test commands, conventions that differ from the defaults, and hard constraints. Move rarely needed procedures into skills, which load on demand, or into docs the agent reads when relevant. The planner flags long files and long skill descriptions with trim suggestions so you can see which sections cost the most.
As published in November 2026: Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 each have a 1M-token context window; GPT-5.5 has 1,050,000 tokens (prompts over 272K input tokens are priced higher); Gemini 3.1 Pro has a 1M-token window. These change often, so the reference table is editable and the provider documentation is authoritative.
Your instructions file and every tool/function schema are counted as always-on, because a running agent has them in context on every turn regardless of what the user asks. For skills, only the frontmatter — the short name and description used for discovery — is always-on; the full skill body is on-demand because progressive-disclosure designs like Claude's Skills only load it once the skill actually fires. This distinction is the reason a generic prompt-cost calculator gives you the wrong picture of a harness: it can't tell always-on tokens from tokens that are free until invoked.
The line-count and description-length thresholds mirror the guidance in our own AI Harness article: instructions files see diminishing returns past roughly 20-30 concise lines, and skill descriptions rarely improve discovery once they run past about 50 words while still costing tokens on every turn they're loaded. These are reasonable starting thresholds, not hard rules — a genuinely complex project may need more, and the flag is a prompt to double-check, not a verdict.
Token counts use the widely cited approximation of about 4 characters per token for English text. This is a rule of thumb, not an exact tokenizer — different models segment text differently, code and non-English text tokenize less efficiently than this ratio suggests, and the only fully accurate count comes from a model provider's own tokenizer. Use these numbers to compare proportions and catch obvious bloat, not as an exact budget.