Your coding agent is reading junk files.
You’re hitting your plan’s limit faster for it.
Your coding agent works from the files on your disk, and a lot of them are things it never needs to read: lockfiles, build output, vendored dependencies. Every token they add counts against your daily and weekly limits the same as your real code — and less junk in context means fewer chances to lose the thread mid-session. ContextPruner writes the config files that tell your agent what to skip, checks the ones you already have against your repo, and keeps all nine current on every push.
No signup needed. Your file list stays in your browser — we never see it.
Paste your file tree — get the config files that tell your agent what to skip.
Paste manifest
Paste your repo’s file manifest to watch the triage run — and see what the junk is costing you.
Paste a file tree and an existing config to grade it — missing rules, dead rules, drift, and what the gaps are costing you.
How the pieces fit together
ContextPruner covers the whole path — generating the configs, checking them, enforcing four of them at the harness level, filtering junk from your agent's searches at runtime, and keeping all nine current. No competing tool covers that full loop.
- Free
Generate
Paste your file tree and get all nine config files in your browser. Your file list never leaves the page.
- Free on the web
Lint
Already have an AGENTS.md or a .cursorignore? Check it against your repo — a graded report shows what's missing, dead, or drifting, with one-click fixes.
- Free · $9/mo
Enforce
The agent's own harness applies four of the nine — .cursorignore, .geminiignore, .codeiumignore, and Claude Code deny rules — a best-effort block, free, with exceptions declared in a committed .contextpruner file. Paid goes deeper: contextpruner serve filters junk out of broad searches at runtime, all on your machine, while a pointed read of a named file always goes through.
- $9/mo
Automate
Your repo changes every day. A GitHub Action or pre-commit hook regenerates all nine on every push, so the configs keep up — your own runner commits, and we never get access to your repo.
- With your key
Watch it work
Your account page shows what your last sync cut — files scanned, tokens kept out of context, what that is worth per month — plus how many syncs have run.
Works with your whole toolchain
One paste generates the exact instruction file every major coding agent reads — including the cross-agent AGENTS.md standard that 19 of them support natively. Four harnesses also apply a generated ignore file at the tool level, marked Enforced ✓ below.
Show all 24 agents ▾Hide the agent list ▴
Claude Code
CLAUDE.mdDoes not read AGENTS.md; bridge with an `@AGENTS.md` import inside CLAUDE.md or a symlink. Also loads `.claude/rules/*.md` alongside CLAUDE.md.
Docs ↗Gemini CLI
GEMINI.mdReads AGENTS.md only via `{"context": {"fileName": "AGENTS.md"}}` in .gemini/settings.json.
Docs ↗Devin Desktop (Windsurf)
.devin/rules/*.mdAider
CONVENTIONS.mdAdd `read: AGENTS.md` to .aider.conf.yml to auto-load rules.
Docs ↗Lovable
AGENTS.mdAlso reads CLAUDE.md and in-app Knowledge settings.
Docs ↗Replit Agent
replit.mdReplit Agent maintains replit.md itself — paste rules into it or use in-app instructions.
Docs ↗v0
(in-app Instructions)No repo instruction file — paste rules into v0's project Instructions.
Docs ↗Pricing shown is indicative, as of July 2026 — check each vendor's docs for current tiers.
The quota drain
Every request your agent makes carries context, and a lot of it is files it never needed. Whether your platform counts tokens, credits, or fast requests, junk reads spend them the same as your real code — you hit the cap sooner and get less done before the meter stops you. Tell the agent what to skip and you stop paying for files that never mattered.
The 70,000-token lockfile
A committed lockfile like package-lock.json can be around 70,000 tokens the model learns nothing from, and it gets pulled in whenever the agent looks at dependencies or debugs an install. The 'prompt is too long' error is what a repo full of files like this looks like to a model — the config tells your agent to skip them before it hits the wall.
Week-six repo sprawl
A project that starts at 40 files is thousands by week six: build caches, abandoned folders, vendored copies sitting in your working directory on disk. When a long session starts to drift, a crowded context is part of the story — less junk in it means fewer chances to lose the thread. Paste your file tree once and the config points the agent back at your real code.
Guides
Practical guides on the config files that steer your coding agent — what goes in an AGENTS.md or CLAUDE.md, and how to keep your context lean.
- What is AGENTS.md?
AGENTS.md is a plain-Markdown file that tells AI coding agents how to work in your repo — what to read, what to skip, how to build. Format and example inside.
- CLAUDE.md vs AGENTS.md: which one do you need?
CLAUDE.md and AGENTS.md do the same job for different agents. Here's which tools read which file, whether you need both, and how to keep them in sync.
- How to reduce your AI coding agent's context
Lockfiles, build output, and vendored dependencies quietly fill your agent's context window and burn your plan's limits. Here's how to cut them out.
- How to write a CLAUDE.md file (with examples)
CLAUDE.md tells Claude Code how to work in your repo. Here's what to put in it, a copy-paste template, and the mistakes that make Claude ignore it.
- AGENTS.md examples and a copy-paste template
Real AGENTS.md examples for common stacks, plus a template you can drop into any repo and adapt in a couple of minutes.
- Cursor rules explained: .cursor/rules with examples
How Cursor's project rules work — the .cursor/rules format, when each rule applies, example rules, and how to stop Cursor reading files it shouldn't.
- Tools to cut your AI coding agent's context, compared
A fair, disclosed comparison of the tools for cutting AI coding context — config generators, linters, and runtime compressors — and how to choose between them.
- Why does my AI keep losing context?
AI coding agents lose context and hit “prompt is too long” when their window fills with junk files. Here’s what’s actually happening — and how to cut it.
- How does ContextPruner work?
How ContextPruner keeps AI coding agents on your real code: it decides what to skip from your file paths and sizes alone — never their contents.
- What is CLAUDE.local.md?
CLAUDE.local.md is your personal, gitignored version of CLAUDE.md — machine-local Claude Code instructions that never reach the team. When and how to use it.
Free where it builds trust. Paid where it saves time.
Free
Generate and lint your config, in your browser.
- Unlimited manifest triage — 100% client-side
- Nine config files: five advisory (AGENTS.md, CLAUDE.md, GEMINI.md, Cursor .mdc rules, copilot-instructions.md) plus four enforced (.cursorignore, .geminiignore, .codeiumignore, Claude Code settings deny rules)
- Lint an existing config against your file tree — graded report + one-click fixes
- Estimated monthly savings counter with live model pricing
- No signup, no upload, no telemetry on your file list
Pro
Your repo changes every day — the configs keep up with it.
- Drop-in GitHub Action + git pre-commit hook
- All nine config files regenerate on every push — advisory and enforced alike
- contextpruner lint CLI — check configs locally or fail CI on drift
- contextpruner serve — filter junk out of your agent's search results at runtime, all local (shell shims + MCP tools)
- One API key unlocks the automation, the lint CLI, and the runtime filter
- We never get access to your repo — your own runner makes the commits
- Usage dashboard on your account page: syncs run, tokens kept out, what it's worth