AgentDiet helps developers measure, analyze, and optimize AI coding agent context consumption. Identify unnecessary overhead, eliminate duplicated instructions, and keep your AI development workflow efficient.
$go install github.com/qansaid/agentdiet@latestzero dependenciesno API keyno network callsno config file
discoveranalyzeoptimizeenforceThe simulated acme-platform repository loads 6,770 always-on tokens across four instruction files on every request, 3.39% of a 200,000-token window, which is grade C against an 8,000-token budget and an estimated $20.31 per 1,000 requests at $3.00 per million input tokens, assuming one request per session. Seven findings include 2,389 tokens duplicated between AGENTS.md and CLAUDE.md.
Numbers come from the simulated example repository, not a benchmark of your codebase.
Reads the files loaded by
by file convention · no runtime integration
Every agent runtime loads a set of instruction files unconditionally: AGENTS.md, CLAUDE.md, .cursorrules, copilot-instructions.md. Nobody measures them. They grow by accretion until the agent is spending thousands of tokens on instructions before it has read your question.
paid per request
not per task
Injected on every request, before you type
Keep it ruthless. Rules only.
Loaded only while editing that subtree
Fine, proportional to the work.
Costs nothing until the agent opens it
Let it grow. Reference belongs here.
Declares tools and servers, not text
Inventory the declarations; do not score their schema.
A tool that reports “your repo has 90k tokens of agent config” collapses these rows and cannot tell you what to do. Only the first one is a bill; the rest is inventory.
Field report: 80 public repositoriesSampled through GitHub code search, not a random population. Token counts are estimates, and duplication is counted per repository, not per session.Each one is ordinary engineering hygiene until you count it.
A rule per incident, a table per architecture decision. Nothing is ever deleted, because removing a rule feels riskier than adding one, so the file only grows.
AGENTS.md · 285 lines · 3,086 tokens
The same rules under a second filename, kept in sync by hand. A runtime may load both, one, or neither, so the duplicated text is quantified rather than a guaranteed per-request charge. The drift is immediate and certain.
AGENTS.md ↔ CLAUDE.md · 2,389 tokens repeated
Schemas, endpoint catalogs and glossaries pasted into the root file. They are reference material, needed for one task in fifty, loaded for all fifty.
one always-on file above 3,000 tokens
On-call escalation paths, incident history, retention policy. All true, all useful, none of it relevant to the function you asked the agent to change.
scoped rules promoted to always-on, silently
Each connected server contributes tool schemas to the prompt as soon as a session opens. Six servers declared for one repo means six surfaces competing for attention. The audit inventories the declarations; it does not measure the schema tokens they add.
6 servers declared · inventory only
The same paragraph in two files is duplication the agent cannot deduplicate. Whether a runtime loads both depends on the agent; the audit quantifies the duplicated content, and the copies drift apart the first time someone edits only one.
## Git workflow
Branch names are <type>/<issue>-<slug>.
Rebase onto main before opening a PR.
Squash-merge is the default.
## Git workflow
Branch names are <type>/<issue>-<slug>.
Rebase onto main before opening a PR.
Squash-merge is the default.
2,389 duplicated tokens identified. Actual savings depend on which files the runtime loads.
The cost line uses $3.00 per 1M input tokens as a default assumption. Change it with --price. It prices the context, not your invoice. The cost assumes one request per session; the CLI reports the same figure per 1,000 sessions, which the site labels as Estimated cost / 1K requests.
AgentDiet walks the repository, keeps the four tiers apart, and turns the always-on number into something you can argue about, budget for, and gate in CI. It reads files and prints a report and nothing else.
Walks the tree, classifies each recognized instruction file into one of four tiers, and skips node_modules, vendor, dist, build and other dependency trees so a vendored AGENTS.md never counts as your prompt.
internal/audit/scan.go
Prose at 4.5 characters per token, fenced code at 3.4, each segment rounded up. Calibrated against cl100k_base on 137 real agent-config files: 8.1% mean absolute error, 87% within ±15%.
internal/audit/estimate.go
Normalizes paragraphs and reports text that appears in two or more files, quantifying the duplicated content. Actual savings depend on which files the target runtime loads. The CLAUDE.md-is-a-stale-copy drift is quantified, and the drift itself is certain.
internal/audit/report.go
Flags one instruction file that is at least 60% contained in another, which is what a copy that was edited once and never again looks like.
blockOverlap()
Any single always-on file above 3,000 tokens is called out by name, because one file that large is reading material, not a rule set.
detectFindings()
Builds an MCP configuration inventory from the config dialects in the wild (.mcp.json, .cursor/mcp.json and .vscode/mcp.json). It identifies declared server names. It does not measure the schema tokens those servers add, because that depends on the runtime and the tools it connects.
mcpServerNames()
Exits 1 when always-on context outgrows --max-tokens. Machine-checkable, so the number can only shrink on purpose instead of by hope.
main.go
A composite action installs the binary, runs the audit, exposes always-on-tokens and grade as step outputs, and fails the job when the budget is exceeded.
action.yml
It estimates, and prints the method in every report. A 2 MB vocabulary would buy precision nobody acts on and still be wrong for the next model.
It never rewrites your instruction files. Findings carry a fix as text; the diff is yours, reviewed by you.
No account, no upload, no telemetry, no network call. One binary, one pass over your working tree.
Unshipped work is listed as unchecked in the repository’s roadmap: --since, .agentdietignore, exact tokenizer mode and SARIF are planned, not features.
Four steps, one pass, no server round-trip. The first two are a command; the last two are decisions you make with the output. Each stage below is interactive.
The walker finds instruction files by convention: no config file, no manifest to maintain. Each hit is sorted into always-on, scoped, on-demand or config, because only one of those four tiers is a bill.
No configuration. Convention beats a manifest you forget to update.
see the full reportA simulated Go monorepo with the drift almost every mature repository accumulates: a 285-line AGENTS.md, a CLAUDE.md that is a full copy of it, per-client rule files restating the same paragraphs, and six MCP servers declared for one service.
simulated fixture
not a benchmark
Always-on tokens
per request, before you type
Grade
8,000 → 1,500 token budget
Estimated cost / 1K requests
at $3.00 per 1M, one request per session
89.5% less always-on context in this fixture
Reference material was reorganized into the right tiers, while unused MCP servers were removed.
This repository is synthetic. The before and after figures describe that fixture only. They are not a benchmark, not a survey, and not a promised saving for your codebase. AgentDiet does not measure agent output quality, before or after.
The CLI prints its cost line per 1,000 sessions because it assumes one request per session. The labels here call the same figure an estimated cost per 1,000 requests; multiply it by the requests a session actually makes if your workflow differs.
MCP findings in that output are an MCP configuration inventory. AgentDiet identifies declared server names; it does not measure how many schema tokens those servers add, because that depends on the runtime and the tools it connects.
The percentage is a consequence of how much duplication the fixture contains. Your repository may have almost none, in which case the report will tell you that too, and the honest next step is to run it on your own tree, which takes one command and sends nothing anywhere.
Run it on your repositoryNo config file to write, no account to create, no daemon to keep alive. The whole interface is a command with six flags and three exit codes.
Audit the current repository. Pass a path to audit another tree.
--jsonoffEmit the full report as JSON with stable keys, safe for dashboards.--badgeoffEmit a shields.io endpoint document instead of the report.--max-tokens0 (off)CI gate: exit 1 when always-on tokens exceed this number.--window200000Context window used for the share column.--price3.00USD per 1M input tokens, used for the cost line.--top8Maximum files listed per section.--versionn/aPrint the version and exit.identical to `agentdiet --help`, which the binary prints itself.
A number nobody checks is a number that grows. Wire the audit into the workflow you already have, and the always-on tax becomes a pull request check that has to be justified.
name: agent-contexton: [push, pull_request]jobs: agentdiet: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 - uses: qansaid/agentdiet@v1 with: max-tokens: 4000Pins @v1, the moving major tag the maintainers update on each release. Pin a full semver tag if you prefer immutable refs.
The check runs on every PR. A change that inflates the always-on tax fails like any other broken invariant, so the reviewer sees it before the merge, not a quarter later.
The action exposes always-on-tokens and grade as step outputs, and --json gives you every recognized file, finding and total if you want your own dashboard instead.
There is no API key to store, no service to reach and nothing to allowlist. The binary runs in the runner, reads the checkout and exits.
AgentDiet is developed in the open under Qansa, licensed MIT, and deliberately small: a stdlib-only Go binary with no runtime dependencies, no network calls and no telemetry. There is no paid tier and no hosted version to upsell you into.
MIT
Use it, fork it, ship it inside a commercial product. The only obligation is attribution.
Open
Design questions, new file conventions and estimator accuracy are discussed in the open.
Templates
Three defect forms: a file that was not detected, a finding that looks wrong, and a bug. Questions and ideas go to Discussions.
Welcome
New conventions and new findings are the most useful contributions, each with a test fixture.
Maintained by Qansa
AgentDiet is a Qansa open-source project. Product identity stays with AgentDiet; the repository, releases and roadmap live under the Qansa organization on GitHub.
Answers describe what the code does today. Where the tool is approximate, it says so.
Something missing? Open a discussion. The question that is asked twice becomes a section here.
Ask on GitHubOne command, no account, no upload. If your context is already lean the report will say so in one line, which is the cheapest possible answer to a question most teams never ask.
$go install github.com/qansaid/agentdiet@latestMIT · Go 1.22+ · no telemetry · no network calls
AGENTS.md
always-onThe canonical rule set, plus schema tables, a glossary and a runbook that only one task in fifty needs. 100% contained in CLAUDE.md.
always-on
Injected on every request, before you type.
Keep it ruthless: rules only.
duplicate · 2,389 tokens
Shared verbatim with CLAUDE.md. A runtime may load both; the audit quantifies the duplicate.
4 always-on files · 6,770 tokens · 3.39% of the window