MCP tools for coding agents
Tokeasy is ten drop-in tools that cut the round-trips, re-reads and log floods out of Claude Code, Codex, Copilot and Cursor. Same model, same quality — about a fifth off the bill, with every saving labeled for what it is.
2-minute setup · cancel anytime · your code never leaves your machine
The savings ledger, as your agent reports it — every number labeled counted, measured or estimated.
01/The problem
of an agent's tokens go to re-reading code it has already seen this session.
re-ingests the growing transcript — each extra round-trip makes every later one more expensive.
an exact-match miss on whitespace wastes the whole turn — then the agent re-reads and retries.
02/How it works
The same everyday task — find a function across files, change it everywhere, verify nothing broke.
Fewer calls, smaller context — the savings compound every turn.
03/Benchmarks
Five real agent tasks, run five ways: vanilla Claude Code, the open-source token savers Headroom, RTK and Caveman, and Tokeasy. Same task, same model — only the efficiency layer changes. Lower is better.
total suite cost vs vanilla Claude Code
the savings of the next-best tool on the same suite
tasks where Tokeasy had the lowest cost
Costs are the full provider bill to complete each task, end to end. Percentages are against vanilla Claude Code on the same task — note the per-call strippers can land above baseline once the agent starts re-fetching what was stripped from its context. Your workloads will vary.
04/The kit
Drop-in replacements for the expensive habits: exploration chains, full-file dumps, brittle edits, log floods, git chatter. Every result opens with a [now: … UTC] stamp, so agents stop wasting a round-trip asking the clock.
Glob, grep, read and imported-by in a single ranked call — matches, context lines and the import graph in one round-trip.
AST-aware truncation: long bodies stubbed, imports, types and signatures kept. The shape of the file without ingesting 2,000 lines.
Batched multi-file edits with fuzzy matching that tolerates whitespace and quote drift — plus TS/JSON/YAML syntax validation before anything ships.
Test and build output as a bounded digest that keeps every failure — a 100 KB pytest run becomes ~2 KB, with a receipt for what was cut and a way to get any of it back exactly.
Branch, ahead/behind, dirty files, recent commits, diffstat, stash and merge state — the whole session-start chain as a single round-trip.
Databases, command output, system state, cross-session memory, the savings ledger. New tools ship to every subscriber the day they land.
05/Install
Pick your harness and paste. The same license key works across all of them.
claude mcp add tokeasy -e TOKEASY_LICENSE_KEY=tk-... -- npx -y tokeasy~/.codex/config.toml
[mcp_servers.tokeasy]
command = "npx"
args = ["-y", "tokeasy"]
env = { TOKEASY_LICENSE_KEY = "tk-..." }.vscode/mcp.json
{
"servers": {
"tokeasy": {
"command": "npx",
"args": ["-y", "tokeasy"],
"env": { "TOKEASY_LICENSE_KEY": "tk-..." }
}
}
}.cursor/mcp.json
{
"mcpServers": {
"tokeasy": {
"command": "npx",
"args": ["-y", "tokeasy"],
"env": { "TOKEASY_LICENSE_KEY": "tk-..." }
}
}
}openclaw plugins install tokeasyopenclaw.json
{
"plugins": {
"entries": {
"tokeasy": {
"config": { "licenseKey": "tk-..." }
}
}
}
}06/Pricing
Every tool, current and future, across every supported harness.
instant license key · cancel anytime
Checkout by Stripe. By subscribing you agree to the Terms & Privacy Policy.
What you get
Unhappy in your first few weeks? We're reasonable about refunds — team@tokeasy.net.
07/FAQ
Two mechanisms: fewer tool calls (one ranked search instead of a glob → grep → read chain) and smaller payloads (AST-truncated reads instead of full-file dumps). Because every tool call re-ingests the growing transcript, both effects compound — each call you skip makes every later call cheaper too.
Yes. Plan usage limits are token-budgeted under the hood, so every token Tokeasy avoids is quota you get back. Fewer, smaller calls means more tasks before you hit a usage window — sometimes up to twice as much out of the same subscription.
No. Your agent keeps using the exact same model — Tokeasy only changes what flows through tool inputs and outputs. If anything, quality goes up: post-edit syntax validation catches broken TS/JSON/YAML edits before they land.
License checks only. Tokeasy runs entirely on your machine; the validate call sends your license key and an instance hash so we can enforce the device limit. Your code, prompts, file paths and query contents never leave your machine.
Claude Code, Codex CLI, VS Code Copilot, Cursor and Windsurf via MCP, plus a native OpenClaw plugin — and anything else that speaks MCP.
Proxies strip each request after the fact, then report "% tokens removed per call". But the agent doesn't know what was stripped — when it needs that information, it calls the tools again to get it back. Those repeated round-trips are where task cost actually lives. Tokeasy works at the tool layer instead: the agent gets what it needs in fewer, smaller calls to begin with, so there's nothing to re-fetch.
Because we publish what each number is: calls saved are counted, bytes returned are measured, bytes saved are estimated — and the dollar figure is priced from posted list rates under session cache economics, with every assumption a named constant in source you can read and change. No percentage of a number nobody can audit.
Ask your agent to run the code_savings tool. It reports the ledger — calls saved counted, bytes returned measured, bytes saved estimated — each number labeled for what it actually is.