pip install mcptoon

# 30-second proof, on your machine — none of your servers, no API key:
mcptoon demo --quick

# Add any MCP server — one command:
mcptoon add everything --stdio npx -y @modelcontextprotocol/server-everything

# What your agent actually reads (names only — 581 tokens, not 71,929):
mcptoon manifest

Your tools stay yours. mcptoon bundles nothing — it's a remote control, not a runtime. The MCP servers you want, you install yourself, one command each (npm/pip/a URL). The one thing it does ship is a demonstration of itself: mcptoon demo-server is 11 tools written in the standard library, and they enter your agent's context only if you add that server deliberately. Delete mcptoon someday? Your MCP servers keep running on their own — not one goes missing.

Mcptoon is the native decoupling layer for MCP tools. It fixes the twin pain of MCP tool listings eating tokens and every AI agent re-configuring tools on its own. Zero config, out of the box: it auto-scans and unifies the MCP tools of every agent on this machine — Claude Code, Cursor, Codex, scripts, CI — shares tool instances globally, and slashes token overhead.

This isn't just us talking

Those numbers are ours, but "loading every tool schema into context is expensive" is not a claim only we make:

• Anthropic: tool schemas flooding the context window is a real pain — one example drops from 150,000 tokens to 2,000 (a 98.7% saving)
• Firecrawl benchmark: the same task cost 1,365 tokens via CLI vs 44,026 via MCP — 32× (full schema loaded upfront)
• Scalekit benchmark: CLI is 10–32× cheaper and 100% reliable; MCP scores 72%
• MCP-Zero (arXiv:2506.01056): on-demand tool retrieval achieves near-constant cost regardless of tool count
• SEP-1576: an open MCP proposal to cut schema redundancy — the problem is acknowledged upstream

We're not the only ones who measured this. mcptoon is the one you can use today, covering every agent at once.

Up and running in 30 seconds

pip install mcptoon                          # pure stdlib, 155KB, zero dependencies

# Add any MCP server — one command:
mcptoon add everything --stdio npx -y @modelcontextprotocol/server-everything

# See every tool available (names-only by default; 255 tools cost 581 tokens):
mcptoon manifest

# Call a tool (JSON output by default; add --toon to save more):
mcptoon call everything echo '{"message":"hi"}'

What mcptoon demo actually prints

No API key, no MCP server of yours, nothing to configure: it boots the official "everything" reference server, calls one tool, and shows the token math on your screen. This is that output, verbatim from v0.7.11 (Windows, Python 3.12) — only the ASCII banner and the closing star-ask are cut:

$ mcptoon demo --quick

  Starting demo server...
  ✓ Demo server ready

  📊  SAME data, 19% fewer tokens:
              21  →          17   (SLIM)

  Format         Tokens    Savings
  ────────── ────────── ──────────
  JSON               21          -
  TOON               17        19%
  SLIM               17        19%

  Official benchmark: 255 tools, 50 servers, tiktoken cl100k_base:

  Format           Tokens    Savings
  ──────────── ────────── ──────────
  JSON             71,929          -
  TOON             47,438        34%
  SLIM              8,282      88.5%
  Compact             581      99.2%

  Now you can:
  ✓ connect every agent with ONE config   →  mcptoon sync
  ✓ expose ALL servers as ONE stdio server →  mcptoon serve
  ✓ never paste tool schemas again         →  mcptoon manifest --slim

  Schemas are fetched, not injected — you pay for a listing, not for every turn
  255 tools listed once: 581 tokens (71,929 → 581, −99.2%, measured)

Column padding follows your terminal width; the numbers do not. The first table is one live tool call measured on your machine; the second is the repo's committed 255-tool benchmark (docs/tiktoken-benchmarks.md), reproduced identically on every run.

No Node on the machine? mcptoon demo-server is the same proof with nothing to download: an MCP server of 11 standard-library tools, no network, no API key.

Claude Code user? Skip the terminal entirely:

/plugin marketplace add activeing123/mcptoon

The plugin auto-installs the CLI (SessionStart hook), wires the mcptoon serve bridge via .mcp.json, and ships a skill that teaches the agent when to compress. /mcptoon-setup is the manual fallback.

Or let mcptoon auto-discover servers already on your machine:

mcptoon quickstart     # discover + configure + list tools — one command

That's it. No hand-written JSON config. No MCP protocol debugging. No polluted context window. The wheel is 155KB with zero dependencies, and mcptoon itself needs no API key and phones nothing home — $0 in service fees, everything runs on your machine.

The problem

Every MCP agent (Claude Code, Cursor, Codex, …) stuffs every tool's full schema into your context window before doing any work:

50 tools  → 14,113 tokens of schema → a 128K context: 11% gone
255 tools → 71,929 tokens of schema → a 128K context: 56% gone

So you unload servers you aren't using and reload them when you are. Back and forth. Adding one new server still means hand-writing a JSON config — one missing comma and everything breaks.

mcptoon fixes this. Your MCP servers stay configured, but their schemas never enter the agent's context by default. The agent just runs mcptoon commands, and only the compact result you asked for enters context — the name index weighs 581 tokens (114 for 50 tools, −99.2%).

Without mcptoon: 255 tools → 71,929 tokens, over half the window
With mcptoon:    255 tools → 581 tokens. 99.2% saved.

Both rows are measured configurations, not one number scaled up and down (tiktoken cl100k_base, assets/benchmark_tiktoken.json). Your mix will differ — compute your own numbers in the browser, 30 seconds, nothing uploaded.

The industry validated the problem — then gated the fix

Token-heavy tool context is no longer a niche complaint — it is now an official engineering problem, and the same answer keeps appearing on every roadmap:

• Anthropic: Tool Search Tool defers tool definitions until needed; Programmatic Tool Calling moves orchestration into code. Both are Claude-platform betas.
• MuleSoft: MCP Payload Optimization productizes clean → distill → compress (its compression stage is TOON). Enterprise gateway only, MCP up to 2025-06-18.

The direction is settled. mcptoon is the version of this answer you can run today, on every agent at once — the results-side discipline without the platform or gateway toll.

Installing MCP servers — one command each

# Install from npm (most MCP servers live here):
mcptoon install brave-search --npm @modelcontextprotocol/server-brave-search

# Install from pip:
mcptoon install my-tool --pip mcp-my-tool

# HTTP/SSE servers:
mcptoon install remote-api --url https://example.com/mcp

# List installed:
mcptoon install --list

# Uninstall:
mcptoon install --remove brave-search

mcptoon connects, discovers tools, generates the handler, registers it. No restart needed. Each install adds 0 KB to mcptoon itself — the CLI stays 155KB with zero dependencies, because servers are external processes your machine runs directly, not code bundled into mcptoon. Four steps, one command, no agent restart.

Any MCP server works:

mcptoon add my-server --stdio npx -y @any/mcp-package
mcptoon manifest    # usable immediately

Install as an agent skill (works with 80+ agents)

Teach your agent to use mcptoon through the open agent-skills ecosystem — the skill is picked up by Claude Code, Cursor, Codex, Cline, Windsurf and 75 more:

npx skills add https://github.com/activeing123/mcptoon --skill mcptoon

Prefer a GUI? ToonDeck

Don't want to hand-edit configs? ToonDeck is a local console for mcptoon: every MCP server and tool in one place with a real health check, one skill folder synced to all your agents, agent launching with live logs, and API keys stored in your OS keychain — never in a plaintext file.

pip install toondeck    # ships the web UI inside the wheel — no node, no build

Pre-alpha; free (Apache-2.0). ToonDeck drives the engine — mcptoon stays the single source of truth underneath.

Works with every AI agent

mcptoon is a CLI. If your agent can run a shell command, it can use mcptoon. No plugins, no SDK, no per-agent setup.

Configure once in ~/.mcptoon/config.json; every agent that can run shell commands shares the same servers and tools. GUI agents that can't? mcptoon sync writes native JSON into each one's own location.

export MCPTOON_AGENT_TYPE=claude   # call results auto-select --toon
# export MCPTOON_AGENT_TYPE=openai   # or keep the default JSON

Your AI can even add tools by itself — no human in the loop:

# Agent needs GitHub access mid-task? It just runs:
mcptoon add github --stdio npx -y @modelcontextprotocol/server-github
mcptoon call github search_repos '{"query":"mcp"}'
# Done. No JSON editing. No restart. No lost context.

The numbers

mcptoon's token savings are two separate bills — know which one you're reading before comparing numbers. The short version: at 255 tools, native discovery costs 71,929 tokens — over half of a 128K context — while the same toolset reads back at 581 tokens through the name index, a 99.2% cut. On the results side, --toon saves 34.0–34.2% versus JSON across the measured set. Both rows are measured configurations (tiktoken cl100k_base, assets/benchmark_tiktoken.json), not scaled estimates.

Bill 1 · Tool discovery (manifest): 99.2% saved by default

This bill comes due before your agent decides "which tool do I use". Native MCP shoves every tool's full schema into context (50 tools: 14,113 tokens; 255 tools: 71,929 tokens). mcptoon sends only the name index — that's where "114, not 14,113" comes from.

Zero action, on by default: mcptoon manifest with no flags is this tier. Want more per tool? --slim (names + parameter types) costs 8,282 tokens (−88.5%); --full adds descriptions on top; --json (full schema) is the baseline.

We measured both rows ourselves — not one number scaled up and down (tiktoken cl100k_base, assets/benchmark_tiktoken.json). Your mix will differ — compute your own numbers in the browser, 30 seconds, nothing uploaded.

Bill 2 · Call results (call): optional, --toon saves ~34%

This bill comes due after a tool returns its result to your agent. mcptoon call outputs JSON by default — yes, the default saves nothing. To shrink results too, add --toon (structured encoding, reversible):

Default:  mcptoon call fetch fetch '{"url":"https://example.com"}'   → JSON (baseline)
Leaner:   mcptoon call fetch fetch '{"url":"https://example.com"}' --toon   → ~34% saved

34% is the measured toon_save value in assets/benchmark_tiktoken.json (34.0–34.2%), not a marketing number.

One line to remember: 99.2% is what you save seeing which tools exist; 34% is what you can further save on results.

Side-by-side (Bill 1, made visible)

One tool's schema costs 37 tokens as native JSON but only 2 tokens as an mcptoon name-index entry — a 95% cut on a single tool. We measured it with tiktoken (cl100k_base).

Without mcptoon (what every MCP client stuffs into context — 37 tokens, measured with tiktoken):

[{"name":"search_web","description":"Search the web for information",
"inputSchema":{"type":"object","properties":{"query":{"type":"string","description":"Search query"}}}}]

With mcptoon (2 tokens):

search_web

With mcptoon --full (6 tokens, parameter info included):

search_web|query:s*

Security

Three layers, all built in:

• No telemetry. No analytics, no crash reports, no call-home.
• No stored credentials. API keys pass straight from your config or environment.
• No dependencies. Pure Python standard library. Nothing in the supply chain to audit.
• No daemon. Pure CLI — no resident process, no listening port, no attack surface.

All commands

mcptoon quickstart              # one-shot start (discover + configure + list tools)
mcptoon list                    # show configured servers
mcptoon manifest                # all tool names (compact by default; 255 tools = 581 tokens)
mcptoon manifest --slim         # names + param types (8,282 vs 71,929 = −88.5%)
mcptoon manifest --compact      # names only (581 vs 71,929 = −99.2%, measured)
mcptoon inspect <server> <tool> # inspect one tool's schema
mcptoon search <query>          # search tools across servers
mcptoon call <server> <tool> '{"args":"here"}'   # call a tool
mcptoon call --auto <tool> '{"args":"here"}'     # auto-find the server
mcptoon call <server> <tool> --stdin             # read large args from stdin
mcptoon add <name> --stdio|--http <cmd|url>     # add any MCP server
mcptoon remove <name>           # remove a server
mcptoon install <name> --npm|--pip|--url <pkg>  # install + auto-generate handler
mcptoon install --list          # list installed
mcptoon install --remove <name> # uninstall
mcptoon sync                    # sync native config to every detected agent
mcptoon plugin install <dir>    # install an Agent Plugins 1.0.0 plugin
mcptoon serve                   # run as an MCP server (stdio/HTTP) — MCP 2026-07-28: stateless-first, server/discover, cacheable list results
mcptoon demo                    # one command, live demo on your machine
mcptoon doctor                  # self-check: Python, config, connectivity
mcptoon usage                   # local call statistics
mcptoon completion ps           # shell completion (bash/zsh/fish/powershell)

Format family: four tiers, compact by default

Discovery and call results each have a set of formats — all optional, and the default is already the leanest tier:

manifest (discovery): compact by default, upgrade only if you want more

Why compact by default, not full? Deciding "which tool do I use" only needs names (581 tokens for 255 tools); parameter details matter at call time, fetched on demand via inspect or manifest --full. Defaulting to full schemas hands the 99.2% right back.

The recommended loop (and why it's load-bearing). Compact names are a catalog, not a calling contract. Measured on 41 live tools (full-schema gold standard, one inference model per condition): an agent that guesses arguments from names alone lands 4/41 ≈ 10% valid calls — the killers are non-guessable names like account and pageId — while an agent that runs inspect <server> <tool> once for the 2–3 tools a turn actually uses hits 41/41 ≈ 100%, identical to injecting every schema. A turn touches a handful of tools, so a few on-demand inspect calls stay far below the cost of a full-schema dump: you keep ~99% of the token savings and full call accuracy. The rule for agents: use manifest to choose, inspect before you call.

call (results): JSON by default, --toon to save

Where these formats come from

• compact: a name list — any tool manager can do it; nothing proprietary.
• slim (name|param:type* signatures): an mcptoon original, implemented in output.py (slim_toon, Apache 2.0); noted in NOTICE.
• full: full JSON Schema — what MCP speaks natively.
• toon (result encoding): integration of the open TOON standard (toon-format/toon v4.1, MIT), vendored from python-toon and credited in NOTICE — not our invention, and we don't claim it.

Do custom formats break MCP compatibility? — No, for three reasons

"Proprietary format = compatibility bomb" is a fair worry. It doesn't apply here:

1 · The protocol layer is always standard JSON-RPC; formats live only in the presentation layer. mcptoon speaks standard MCP to servers (initialize / tools/list / tools/call — and since the 2026-07-28 GA bridge, server/discover and stateless requests too; the initialize handshake remains for legacy clients). compact/slim/toon only affect the "mcptoon → agent" output rendering — not a single byte toward the server. Servers always see standard JSON; they don't even know these formats exist.

2 · --toon isn't proprietary; it's an open standard. TOON (Token-Oriented Object Notation) is an external open standard (toon-format/toon v4.1, MIT, official TypeScript reference implementation). We integrate python-toon (MIT); tests/test_toon_cross_validate.py verifies decode(encode(x)) == x case by case.

3 · There's a fallback; worst case you fall back to JSON. If --toon decoding fails it falls back to JSON automatically (--fallback-json); and call results are JSON by default anyway — --toon is optional. Want the full schema back? One --full is native MCP. No lock-in.

In one line: zero protocol changes, formats live in the output layer, worst case falls back to JSON. The feared "server can't understand the custom format" can't happen — servers always hear standard JSON-RPC.

How it works

mcptoon is a CLI tool, not an MCP client library. Your agent doesn't connect to MCP servers — it runs mcptoon commands. Schemas live on disk in ~/.mcptoon/config.json, out of the context window by default.

Two-layer decoupling:

Layer 1: mcptoon CLI (155KB, zero deps)
         runs in the agent's shell. schemas stay out of context by default.
                    │
Layer 2: the actual MCP servers (npm/pip packages)
         start only when a tool is called. Zero cost when idle.

• 1,000 servers configured → 0 running, until you call one
• mcptoon bundles nothing — you add what you want, one command each (the sole thing it ships is mcptoon demo-server, a self-demo you opt into)
• Delete mcptoon? Your MCP servers keep running independently

Why a CLI, not a proxy

MCP's premise: every capability is a server, and your agent must be configured to reach it. That premise is why one new tool means editing per-agent JSON in a different format for each, restarting everything — and why every agent re-pays the full schema cost before doing anything.

A command line is the one interface every agent already has. And the form factor is measurably cheaper, independent of anything mcptoon does:

• Firecrawl's benchmark: the same task cost 1,365 tokens via CLI vs 44,026 via MCP — 32×
• Scalekit's benchmark: CLI 10–32× cheaper, 100% reliable vs MCP's 72%

If you truly need the proxy form, mcptoon serve is that mode — all configured servers behind one MCP endpoint, with connection pooling and per-agent API keys.

Contributing

git clone https://github.com/activeing123/mcptoon.git
cd mcptoon
pip install -e . --no-build-isolation
pip install pytest pytest-cov
python -m pytest tests/ -v   # 925 passed, 1 skipped

Zero dependencies is a hard rule — our test suite gates every change (915 tests green before merge). See CONTRIBUTING.md and DEVELOPERS.md.

The codebase: 13,415 lines of Python across 23 modules, zero third-party dependencies.

Ecosystem

• ToonDeck — GUI console for mcptoon (pre-alpha): every MCP server, tool, model and API key in one desktop deck, with mcptoon as its engine. Prefer pointing and clicking over typing commands? Same engine, graphical.