tool-tax

Find the token tax hiding in your MCP config.

tool-tax doctor reads Claude Code, Cursor, VS Code, and Cline MCP configs, probes stdio servers with tools/list, and shows how much tool-schema text your agent may load before the user asks a question. It also flags risky config shapes such as literal secrets, shell-eval commands, unpinned package runners, and broad filesystem scopes.

It also scans JSON/YAML tool manifests and OpenAPI files, diffs catalog changes in CI, writes a slim progressive-loading index, and can test whether a lazy-schema stdio proxy helps your host.

pipx install tool-tax

tool-tax doctor --mcp-config .mcp.json
tool-tax doctor --include-global --no-probe
tool-tax doctor --mcp-config .mcp.json --no-probe --fail-on-risk-level high
tool-tax scan examples
tool-tax diff old-tools.json new-tools.json
tool-tax pack examples --out .tool-tax

Example result:

Servers: 1
Tools: 14
Total tool tax: 2,102 est. tokens
Slim index: 647 est. tokens
Savings: 1,455 est. tokens (69.2%)

Why

Agents keep getting more tools. MCP servers, browser tools, GitHub tools, database tools, and internal APIs all ship long schemas. If every schema is loaded up front, your agent pays a context tax before it starts working.

tool-tax gives that tax a number and lets CI enforce a budget.

Install If

• You use Claude Code, Cursor, or another agent with multiple MCP servers.
• You expose internal OpenAPI routes or tool catalogs to an agent.
• You want CI to catch tool-schema bloat before it lands in a pull request.
• You need evidence before claiming that a lazy proxy saves runtime context.

Do Not Use If

• You need a provider billing meter. est. tokens are local estimates.
• You need a full MCP gateway for resources, prompts, streamable HTTP, or auth.
• You are trying to compress normal chat prompts or model responses.

What It Does

• Finds tool definitions in JSON, YAML, and OpenAPI files.
• Inspects project and user MCP config files for Claude Code, Cursor, VS Code, and Cline.
• Probes live MCP stdio servers through initialize + tools/list.
• Flags MCP config risks before probing a server.
• Estimates token cost for each tool schema.
• Ranks the most expensive tools.
• Diffs base/head catalogs for PR budget checks.
• Generates a slim tool-index.json plus separate schema files.
• Runs a stdio proxy with list_tools, get_schema, and call_tool wrappers.
• Fails CI when the tool catalog grows past a budget.
• Posts or updates a GitHub PR report comment when configured.

Install

From PyPI:

pipx install tool-tax

From GitHub:

pipx install git+https://github.com/SihyeonJeon/tool-tax.git@v0.9.0

From a clone:

git clone https://github.com/SihyeonJeon/tool-tax.git
cd tool-tax
python3 -m pip install -e .

Installs one runtime dependency: PyYAML.

Use

Inspect an MCP config:

tool-tax doctor --mcp-config .mcp.json
tool-tax doctor --mcp-config .cursor/mcp.json --format json
tool-tax doctor --include-global --no-probe
tool-tax doctor --mcp-config .mcp.json --no-probe --fail-on-risk-level medium

Scan a repo:

tool-tax scan .

Scan a live MCP stdio server:

tool-tax mcp -- npx -y @modelcontextprotocol/server-filesystem /tmp
tool-tax mcp --pack-out .tool-tax-mcp -- npx -y @modelcontextprotocol/server-memory

Run the lazy-schema proxy:

tool-tax proxy -- npx -y @modelcontextprotocol/server-filesystem /tmp
tool-tax proxy --call-timeout 120 -- npx -y @modelcontextprotocol/server-filesystem /tmp

Write Markdown and JSON reports:

tool-tax scan examples --format md --out tool-tax-report.md
tool-tax scan examples --format json --out tool-tax-report.json

Create a progressive-loading pack:

tool-tax pack examples --out .tool-tax

That writes:

.tool-tax/
  tool-index.json       # small always-load index
  schemas/*.json        # full schemas loaded only when needed

Fail CI on tool bloat:

tool-tax scan mcp-tools.json --max-tokens 12000 --max-tool-tokens 750

Compare a pull request:

tool-tax diff base-tools.json head-tools.json --max-delta-tokens 500

Slice a large OpenAPI file before scanning or packing:

tool-tax scan openapi.json --tag payments
tool-tax scan openapi.json --path /v1/payment_intents
tool-tax pack openapi.json --operation "PostPayment*" --out .tool-tax-payments

Use it as a GitHub Action:

- uses: SihyeonJeon/tool-tax@v0.9.0
  with:
    path: .
    max-tokens: "12000"
    max-tool-tokens: "750"

Summarize multiple public scans:

tool-tax benchmark docs/benchmarks/public-catalogs.yml
tool-tax benchmark docs/benchmarks/public-catalogs.yml --format json --out public-catalogs.json

Output

# Tool Tax Report

Grade: **lean**

| Metric | Value |
| --- | ---: |
| Tools | 7 |
| Full tool tax | 1,144 est. tokens |
| Slim index | 309 est. tokens |
| Slim-index savings | 835 est. tokens (73.0%) |
| Worst tool | 255 est. tokens |

Supports

• Live MCP stdio servers
• MCP config files with mcpServers or VS Code-style servers
• Lazy-schema MCP stdio proxy
• MCP-style JSON/YAML tool arrays
• Agent tool manifests with name, description, and inputSchema
• OpenAPI paths operations
• OpenAPI slicing by --tag, --path, and --operation
• Nested JSON catalogs
• GitHub Step Summary and PR comment reports

Benchmarks

These are local estimates for upfront schema text, not provider billing totals. See Estimator for the method and limits.

Public catalog benchmark: 10 catalogs, 3,429 tools, 1,442,056 estimated schema tokens. The benchmark corpus drops to a 169,423-token slim index before exact schemas are fetched.

Config risk lint sample: 2 servers, 5 findings, high risk.

Direct-vs-proxy upfront schema tax:

Host-level benchmarks:

Repo Shape

src/tool_tax/   # library + CLI
tests/          # unittest smoke coverage
examples/       # sample MCP/OpenAPI catalogs and reports
docs/           # usage docs, benchmarks, and public scan reports

What It Is Not

• Not a provider billing meter.
• Not prompt compression for normal chat messages.
• Not a security sandbox for untrusted MCP servers.
• Not a full MCP gateway.
• Not a universal runtime token saver across all agent hosts.

It measures the up-front schema tax, creates a smaller index, catches catalog growth in CI, and includes an experimental stdio proxy that exposes three wrapper tools so upstream schemas can be fetched only when needed. The proxy is intentionally narrow: stdio, tools/list, and tools/call.

More

• Roadmap
• Estimator
• Compared
• Changelog
• Doctor
• Doctor host config benchmark
• Host matrix
• Naive MCP host benchmark
• Claude Code E2E benchmark
• Public scan gallery
• Proxy benchmark

License

MIT