contextcost

What does it cost an AI agent to read your repository — and how much of that is waste?

uvx contextcost .          # no install, no config, results in seconds

The 3-second version

Real output on plotly.js (full public checkout at 71a2ff7, 2026-08-24, contextcost v0.5.0):

$ contextcost plotly.js/

  63,831,059 tokens to read this repository   ±14% estimated, no tokenizer
  2717 text files · 1346 binaries not counted · 6 paths ignored

WHERE IT GOES
  test                   35,583,286  ████████████████████████████  56%
  dist                   24,598,872  ███████████████████·········  39%
  src                     1,596,194  █···························   3%

LARGEST FILES
  4,578,500  test/image/mocks/gl3d_snowden_altered.json  sampled
  4,578,201  test/image/mocks/gl3d_snowden.json  sampled
  2,593,972  dist/plotly-strict.js  sampled

SAVING
  63,831,059 → 37,008,917 tokens   42% saved
  Measured by walking the repository again with the proposal applied,
  not by subtracting what was dropped.

26.8 million tokens of that repository's reading cost is compiled bundles, recorded test fixtures, and generated files — enough to fill a frontier model's 200K-token context window about 134 times, spent mostly on JSON number matrices and minified output. An agent asked to fix a chart bug reads none of it usefully.

And the estimate is trustworthy: re-run with --accurate, the real tokenizer (cl100k_base) counts 63,363,404 tokens — the estimate above landed 0.7% off, well inside its stated ±14% band. On data-heavy repositories that gap is exactly the thing naive character-counting gets wrong by 30–70%.

The "42%" is not a sum of opinions. The tool proposes cuts, then walks the repository a second time with the proposal applied, so the saving is the difference between two measurements — and if a pattern had caught anything it wasn't supposed to, you'd see the narrowing in the report. This scales down too: seven real repos measured ran from 46.5% waste (dask) to 2.7% (astropy) — clean repos correctly get told they're clean.

Point it at any repository. It measures what reading that repository costs in tokens, works out which files are spending that budget without earning it, and then re-measures the result — so every number it reports is one it observed.

Install

# try it on any repo right now, nothing installed:
uvx contextcost .

# or install it:
pip install contextcost

No dependencies. Python 3.9+. Also works with pipx run contextcost ..

Why this exists

Every coding agent — Claude Code, Cursor, Codex, Copilot Workspace, an in-house one — spends part of its context window just working out what is in your repository. That budget is finite and it is charged per token, and most repositories quietly spend a large fraction of it on files that often add little to an ordinary source-code task: lockfiles, minified bundles, vendored dependencies, snapshot fixtures, generated clients.

The first repository this was ever pointed at had 55% of its entire context cost in a single generated CSV.

There are good tools for packing a repository into a prompt — repomix, gitingest, code2prompt, files-to-prompt. This is not one of them. Packing is a solved and crowded problem. Auditing what the packing will cost you, and reducing it with evidence, was not.

What it will not do

Stated up front, because a tool that measures something is only useful if you know where its numbers stop.

It does not use a real tokenizer. An exact count needs tiktoken — a compiled dependency with a wheel per platform. A tool whose pitch is "find out what your repo costs in ten seconds" cannot open with a build toolchain. So it approximates by character class and prints its error bound next to every total.

That bound is measured, not asserted. docs/calibrate.py encodes a corpus with cl100k_base and compares:

The bound the tool actually prints is ±14%: the measured 95th percentile plus 20% headroom. The corpus is this repository's own files, so every commit changes it slightly, and a bound sitting exactly on the measurement would turn ordinary editing into a red build — where the tempting fix is to widen the bound, which is how a number stops meaning anything.

Two caveats that belong here rather than in a footnote. It is one tokenizer — Anthropic and most others do not publish theirs, so this is a proxy, and "byte-pair encoders land close to each other" is doing real work in that sentence. And the corpus is this repository's own files plus synthetic dense and CJK samples; it is real code and real prose, but it is not yours.

CJK is counted per script, not as one thing

Charging Chinese at the Latin rate under-counts it roughly threefold, so CJK has always been counted separately. What was wrong until recently is that it was counted as one category, and the scripts are not close to each other:

Traditional Chinese costs 44% more per character than simplified for the same sentence, because the tokenizer has far fewer merges for it. A single constant under-counted it by 30% — and traditional is what this project's author writes documentation in, so the first real user would have been the one mis-billed.

Simplified and traditional share a Unicode block, so they are told apart by looking for characters that exist only in the traditional set. Measured on prose: 27% of traditional Han characters trip that detector, and 0% of simplified ones.

Numeric data dumps are their own class

Running --accurate against plotly.js exposed the estimator's next blind spot: it read 45 M tokens where the tokenizer said 64 M. The gap was not source code — it was recorded test fixtures and JSON number matrices, files whose bodies are thousands of small integers (gl2d_parcoords_blocks.json was under-counted by 71%). A byte-pair encoder merges digits poorly, so the more of a file is digits, the more tokens per character it costs — the opposite of what the code ratio assumed.

Files that are mostly digits now form a numeric class with its own ratio, derived from the measured digit share. On plotly.js this took the whole-tree error from 29.4% to 1.2%; on astropy from 20.8% to 12.6%; h5py and pandas, which have almost no such files, were unchanged. The class is conservative on purpose — a file qualifies only if it already looked like code, at least 10% of it is digits, letters are under 25%, and the digits outnumber the letters — so ordinary number-heavy source stays on the code path.

For most of this project's life that bound read ±12%, and it had been chosen rather than measured — the comment beside it cited a calibration script that did not exist. When the script was finally written, the true figure was more than four times worse, and fixing the ratios it exposed (source code is 4.14 characters per token, not the 3.15 that had been reasoned out; dense content is bimodal and no single ratio fits it) is what produced the table above. That is recorded in estimate.py rather than quietly corrected, because a tool that argues against unverified numbers should say when it shipped one.

It will not decide the ambiguous cases for you. Findings carry a confidence: certain (the file says what it is, or its name is reserved by the tool that wrote it), likely (a strong path convention), and possible. Confidence describes the evidence for the file's category, not universal irrelevance: a lockfile is machine-written with certainty and can still be essential during a dependency upgrade. Read the proposal in the context of the work you use the agent for. That last tier — mostly large data files — is never excluded automatically, because a large CSV is waste in a web app and is the entire subject in an analysis repository, and nothing visible from the file system tells those apart. Those are listed separately, with the rule's reasoning, for you to judge. --include-possible moves them in.

It never edits your repository unless you ask. The default output is a proposal. --write-ignore targets the selected consumer's native ignore file; --write-gitignore remains available as an explicit compatibility option. Write mode refuses a symbolic-link destination instead of following it beyond the repository boundary.

.contextcostignore — exclusions that belong to the measurement alone. The right exclusion for an AI context budget is often wrong everywhere else: a recorded test fixture should reach an agent and stay tracked by Git. contextcost --emit-ignore writes the verified proposal into this project-local file, which contextcost reads for every consumer (even with --no-gitignore) and applies last, so its patterns win — including ! re-inclusions, if you want to rescue files from a broader rule. Accepting a saving is one command and changes nothing outside this tool's own measurement.

It does not reproduce a live product's prompt or bill. Consumer profiles model documented ignore-file inputs: the set of text files eligible for context. They do not reproduce semantic retrieval, repo maps, compression, tool calls, product default exclusions, a proprietary tokenizer, or how much one particular request actually sends.

It has no users yet. This is a new tool. The estimator's error bound is measured against a reference tokenizer, and the reduction is measured rather than estimated, but neither of those is the same as having been run against a thousand repositories by people who did not write it.

How the saving is verified

This is the part worth being suspicious of in any tool that claims one, so here is the mechanism in full.

1. Walk the repository, respecting the selected consumer's ignore inputs. Attribute a cost to every eligible file.
2. Classify what looks wasteful, with quoted evidence per file.
3. Turn the findings into ignore patterns.
4. Walk the repository again with those patterns applied.
5. Compare the files that disappeared against the files that were proposed. Those two sets must be equal. If a pattern took anything extra — a docs/ rule that also caught docs/guide/writing.md, or a PNG sitting beside a minified bundle — the patterns are narrowed to exact paths, the repository is walked a third time, and the report says the narrowing happened.

Step 5 is the one that matters. A saving computed by adding up what a tool decided to drop cannot tell the difference between a pattern that worked and a pattern that matched too much: the number goes up either way.

On an ordinary small project

The example below is generated by docs/build_docs.py: a small web project with some source, a lockfile, a bundle, a vendored widget and a snapshot file. Nobody would call it bloated.

37,603 tokens to read 16 text files (estimated, ±14% — see below for why there is no tokenizer).

Excluding those leaves 23,788 tokens — a 37% reduction, and that number is the difference between two walks of the repository, not a sum of what was dropped.

Usage

contextcost                       # measure the current directory
contextcost path/to/repo          # measure somewhere else
contextcost --json                # machine-readable, schema v1, for scripts and CI
contextcost --json-schema         # print the --json key contract
contextcost --markdown            # GitHub-flavoured Markdown, for PR comments
contextcost --markdown --badge    # ...with a shields.io badge line for a README
contextcost --accurate            # exact counts via tiktoken (see below)
contextcost --include-possible    # also act on large data files
contextcost --consumer cursor     # include .cursorignore in the measurement
contextcost --consumer aider      # include .aiderignore in the measurement
contextcost --consumer repomix    # include Repomix's documented ignore files
contextcost --consumer cursor --write-ignore  # append to .cursorignore
contextcost --emit-ignore         # append to .contextcostignore (this tool only)
contextcost --write-gitignore     # explicit legacy .gitignore destination
contextcost --no-gitignore        # count files git would hide
contextcost --top 20              # more rows per section
python -m contextcost --version   # module entry point also works

Exact counts: --accurate

The default numbers are estimates with a measured ±14 % error bound, and for most decisions — "is this repo worth reading", "which files are the problem" — that is the right resolution. When a number will be quoted, --accurate counts with the real tokenizer:

pip install 'contextcost[accurate]'
contextcost --accurate

Three things stay true when it is on:

• Both numbers are shown. The estimate keeps its band beside the exact figure, and the report says whether the estimate landed inside it. A tool that prints only exact numbers is asking you to forget it was ever wrong.
• Sampling stays sampling. Files above 2 MB are still extrapolated from a prefix and marked sampled, rather than spending seconds being precise about one big file.
• Zero dependencies stays the default. tiktoken is never imported unless you pass the flag; without it installed, --accurate exits with code 3 and the install hint.

Measured on this repository itself, estimate and exact agree within 1 %. On repositories full of numeric data dumps they can diverge far more — which is exactly the kind of thing running --accurate once will tell you.

Consumer-native ignore files

The same repository has a different eligible file set in different tools. ContextCost therefore measures the selected consumer and writes a verified proposal to the file that consumer actually documents:

All non-generic profiles also use nested .gitignore files unless --no-gitignore is supplied. Cursor documents .cursorignore as the stronger boundary for keeping files out of AI requests; Aider documents .aiderignore for large repositories; Repomix documents .repomixignore in the same syntax as .gitignore. The exact scope and the limitations of each model are recorded in consumer profiles.

This distinction matters for tracked files. Git itself says that .gitignore applies to intentionally untracked files and does not stop tracking a file already in the index. A consumer may still use that pattern as its own context filter, but writing .cursorignore, .aiderignore, or .repomixignore expresses the intended AI-tool boundary directly.

The exit code is 1 when an actionable context-waste candidate was found and 0 when it was not, so this works as a CI check:

- name: Keep the context budget honest
  run: pipx run contextcost --quiet

That exit code will bite you under set -e. "Found something" is not an error, but bash -e cannot tell the difference and will abort your script on it. This tool's own CI failed on exactly that the first time it ran. When you want the output rather than the verdict, say so:

contextcost --json > cost.json || true

To gate on total size rather than on removable waste, --fail-over BUDGET exits 4 when the measured total exceeds BUDGET tokens — with --accurate that is the exact count, otherwise the estimate. Exit code 4 is distinct from 1 on purpose: waste (1) is a cleanup task, over-budget (4) is a scoping decision.

- name: The whole repo must fit an agent's window
  run: pipx run contextcost --fail-over 200000 --quiet

In a pull request

--delta measures what a change does to the context budget: it walks a base checkout and the head tree per file, compares them by path, and classifies the head tree so "+32k of this is a lockfile" is a measurement, not a guess from filenames. The bundled Action posts (and keeps updated) that report as a PR comment:

# .github/workflows/contextcost.yml in any repository
name: contextcost
on: pull_request
jobs:
  delta:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write
    steps:
      - uses: CAOShurong/contextcost/action@main

Or run it by hand. The base can be a second checkout of the same repository — for example the PR's merge base next to the working tree:

contextcost . --delta ../repo-at-merge-base --markdown

— or simply a git revision of the repository being measured, which is exported automatically (no second clone, no worktree):

contextcost . --delta main            # uncommitted changes vs the branch
contextcost . --delta HEAD~1          # what the last commit did to the budget
contextcost repo/ --delta v4.0.0      # a tagged release vs its checkout

## contextcost — context delta

**53,834 → 95,631 tokens** (+41,797, estimated ±14%).

| rule | tokens | share of repo |
| --- | --- | --- |
| **lockfile** | +32,232 | 34% |
| unclassified *(real work?)* | +9,565 | 10% |

As a library

from contextcost.reduce import reduce_repository

result = reduce_repository("path/to/repo", consumer="aider")
print(result.before, "->", result.after)   # both measured
print(result.patterns)                     # what to add to .aiderignore
print(result.deferred)                     # what it refused to decide

walk_repository, classify and reduce_repository are all usable separately, and every dataclass has as_dict().

As an MCP server

Coding agents that speak MCP can call this tool instead of shelling out — same measurement, same schema, one tool call:

contextcost mcp    # line-delimited JSON-RPC on stdio

Two tools are served. estimate(repo) returns the full --json payload; propose(repo) returns the measured exclusion proposal plus a ready-to-paste ignore block. Stdio transport, stdlib-only JSON-RPC 2.0 — no SDK dependency:

{"jsonrpc": "2.0", "id": 1, "method": "tools/call",
 "params": {"name": "estimate", "arguments": {"repo": "/path/to/repo"}}}

Claude Code, Cursor and Codex CLI each take a two-line config — copy-paste blocks live in docs/coding-agents.md. Any MCP client then gets "what does this repo cost and what is wasting it" as a tool call beside the model.

Development

python -m pytest -q                      # 196 tests, no configuration needed
python -m ruff check src tests docs
python docs/build_docs.py                # regenerate the figures and README
python docs/build_docs.py --check        # CI fails if they are stale

The figures above are generated from a real run against a generated example repository, and CI fails if the README's numbers drift from what the code actually produces.

Verify a release

Starting with v0.2.0, every GitHub release includes a SHA256SUMS manifest and GitHub build-provenance attestations for the wheel and source distribution:

sha256sum --check SHA256SUMS
gh attestation verify contextcost-0.2.0-py3-none-any.whl \
  --repo CAOShurong/contextcost

The GitHub and PyPI files are built once in the same release workflow. Verify the downloaded bytes rather than treating a tag or a green job as proof of the artifact you installed.

Licence

MIT.