Contextbudget

contextbudget

contextbudget estimates the token cost of a directory tree for an agent context window. It walks a tree, estimates the cost of every text file, compares the total against a budget, and reports the largest files and the subtrees that were skipped, so you can decide what to exclude before a tree eats the whole window.

The estimator, honestly

The token numbers in this tool are a heuristic estimate, not a tokenizer count. contextbudget decodes each counted file as UTF-8 text and applies one deterministic rule: tokens = ceil(chars / 4), where chars is the length of the decoded text. Real tokenizers split text differently depending on the model, so a real tokenizer will give a different number for the same file, sometimes higher and sometimes lower. Use these numbers to compare files with each other, to rank the biggest offenders, and to check a tree against a rough budget. Do not use them as an exact prediction of what any specific model will charge or accept.

Install

npx @reneturcios/contextbudget budget [path]

Or install it as a dependency of your own tooling and use the MCP server form below. The package requires Node 20 or newer.

CLI usage

contextbudget budget [path] [--json]contextbudget mcpcontextbudget --version

• budget [path]: check the tree at path, which defaults to the current directory. Prints a human readable report, or the full JSON report with --json.
• mcp: run the MCP server over stdio.
• --version: print the version.

Exit codes:

• 0: the verdict is WITHIN-BUDGET.
• 1: any other verdict, currently OVER-BUDGET.
• 2: usage error, or the target does not exist.
• 141: the reader closed the output pipe early (the shell visible equivalent of SIGPIPE).

Budget

The default budget is 200000 estimated tokens. Set the environment variable CONTEXTBUDGET_BUDGET to a positive integer to use a different budget:

CONTEXTBUDGET_BUDGET=50000 contextbudget budget ./src

If the variable is unset, unparsable, or not positive, the default of 200000 applies.

Verdict rules

• The verdict is WITHIN-BUDGET when the total estimated tokens are less than or equal to the budget, and OVER-BUDGET otherwise.
• The walk skips these directories by default and never descends into them: node_modules, .git, dist, coverage, .next, vendor.
• Files over 1 MiB (1048576 bytes) are skipped and not counted.
• Binary files, detected by a NUL byte in the first 4 KiB, are skipped and not counted.
• Symbolic links are not followed.

Findings:

• One high finding, rule large-file, for each counted file whose estimated tokens are over 10 percent of the budget.
• One medium finding, rule skipped-large, for each file skipped for exceeding the 1 MiB limit.
• One info finding, rule skipped-dir, for each directory skipped by the default skip list.

Ordering is deterministic: files are ranked by estimated tokens descending, then path ascending, and directory entries are visited in sorted name order. Reports carry no timestamps.

Report shape

The JSON report has tool, version, target, verdict, counts, findings, and summary. The headline numbers live in counts:

• files: files counted.
• totalTokens: total estimated tokens across counted files.
• totalChars: total characters across counted files.
• budgetTokens: the budget in effect.
• remainingTokens: budget minus total, negative when over budget.
• skippedDirs, skippedBinary, skippedLarge: skip counts.
• findings, findingsHigh, findingsMedium, findingsInfo, findingsLow: finding counts.
• ext:<extension>:files and ext:<extension>:tokens: per extension totals, for example ext:.ts:tokens.
• top:<relative path>: estimated tokens for each of the top 10 largest files.

MCP

The MCP server exposes one tool:

• budget_path: takes a path string and returns the same JSON report as contextbudget budget <path> --json.

Run it with contextbudget mcp, or through the registry entry in server.json.

Limits

• Text files only: binary files and files over 1 MiB are skipped, so their real cost is invisible to the report except as skip counts and, for large files, medium findings.
• The estimator is a heuristic, as described above. It is not a tokenizer and does not know any specific model's vocabulary.
• The skip list is fixed. There is no flag or config file for custom skips or custom budgets per directory in this version; the only budget control is CONTEXTBUDGET_BUDGET.
• Counts reflect file contents only. File names, directory names, and any framing an agent harness adds around file contents are not estimated.

Development

npm installnpm run buildnpm test