This repository publishes about thirteen hundred free calculators and checkers as web pages: a 401(k) loan versus withdrawal calculator, a 1099 vendor CSV checker, roofing bid comparisons, an estimator for GitHub Actions minutes, and so on. Most were built by unattended AI coding agents, with an owner steering a few hours a week.
Many of these tools are more useful to an agent working for a person than to the person. A bookkeeper's assistant checking a vendor export, or a coding agent estimating what a workflow change will cost, needs the numbers, not a form to click through. Until this week the only way for an agent to reach most of them was to drive a browser.
platform/agents/tools.mjs, merged in PR #8672 on 20 October 2026, removes that step. It is one Node file with no dependencies. From a clone of the repository it searches the catalogue, describes how to call a tool, and calls it, printing JSON. The same commands run as a local Model Context Protocol (MCP) server for Claude Code, Claude Desktop, Cursor, VS Code and other MCP clients.
The numbers
I counted these on main at commit 1ad1d16c on 20 October, by running the module's own catalogue() and engines() functions over every venture directory, and ls ventures/*/cli.mjs for the last row.
| Measure | Value |
|---|---|
Venture directories with a web page (index.html) | 1,283 |
| Callable without a browser | 1,262 |
| No engine module to call | 15 |
| Refused by the safety gate | 6 |
Tools that also ship their own cli.mjs | 169 |
The README for platform/agents/ gives the same 1,262 / 15 / 6 split. The test file, tools.test.mjs, ran 15 tests with 15 passes when I ran it. One of them imports every tool's engine and fails if one stops loading outside a browser, for example because it touches document at import time.
How it works
Every tool page here does its arithmetic in plain ES modules sitting beside the page: calc.mjs, engine.mjs, check.mjs and similar. The page's <script type="module"> imports them, reads the form, and calls an exported function.
tools.mjs reads that same script tag to find which modules the page computes with. It leaves out plumbing that is not part of the calculation (usage.mjs, share.mjs, byok.mjs, ai.mjs). Then it imports those modules in Node and calls their exported functions with JSON arguments.
There is no per-tool code. A tool is covered the day it lands, and the numbers match the page's, because they are the page's own functions.
Most engines export a DEFAULTS object holding the page's default inputs. call --input lays your object over DEFAULTS, which is what the page does with its form, so you pass only what differs. Asking the 401(k) calculator about a $20,000 need against a $60,000 balance returns the full argument list it used (tax rates, growth, loan rate and term filled in from the defaults) and a result with a maximum loan of $30,000 and a growth sensitivity table. That is an example of the output shape, not advice: these tools give estimates from the inputs you give, not legal, tax, financial or professional advice.
Using it
You need Node 18 or later and a clone. There is no install step, and nothing needs the network after the clone.
git clone --depth 1 https://github.com/duncan-haywood/oss-monorepo && cd oss-monorepo
node platform/agents/tools.mjs search 401k loan
node platform/agents/tools.mjs describe 401k-loan-vs-withdrawal-calculator
node platform/agents/tools.mjs call 401k-loan-vs-withdrawal-calculator analyze --input '{"need":20000,"balance":60000}'
node platform/agents/tools.mjs call 1099-vendor-csv-checker analyze --stdin < vendors.csv
node platform/agents/tools.mjs run actions-minutes-estimator -- --json .github/workflows| Command | What it does |
|---|---|
search <words> | Every word must match the slug, title or description; slug and title hits rank first. |
describe <slug> | Each engine's exported functions with parameter lists and doc comments, its constants such as DEFAULTS, the live page URL, and the tool's own CLI if it has one. |
call <slug> <function> | Calls one function. --input is an object over DEFAULTS, --args a JSON array, --stdin passes pasted text or CSV as the first argument. |
run <slug> -- <args> | Runs the tool's own cli.mjs, for the 169 that have one. |
mcp | A stdio MCP server with search_tools, describe_tool, call_tool and run_cli. |
The repository root has an .mcp.json, so Claude Code offers the server when you open a clone. Other clients take the same block with an absolute path to tools.mjs and the argument mcp.
The intended loop for an agent is: search in the user's words, describe the best match to see its functions and defaults, then call it with only the inputs that differ.
The safety gate
An MCP server that imports and runs code on request needs a clear rule about what it will load. This one has a static gate and a runtime limit.
The static gate runs before any import. A module loads only if:
- everything it imports, followed recursively, is a sibling file in the tool's own directory or a
node:builtin, with no npm packages and nothing outside the directory; - none of those files imports a network or process module. The blocked list is
http,https,http2,net,tls,dgram,dns,child_process,worker_threads,cluster,vmandinspector; - none of them contains
fetch(,WebSocket,eval(,new Function(orprocess.binding.
This is the same import check the npm MCP bundle uses (platform/distribution/mcp.mjs). describe lists a refused module and the reason, so an agent knows to use the web page instead.
The gate reads source text, so it is deliberately conservative. Of the six tools it refuses, some do go online (a CORS checker, an SRI checker, a PWA installability checker). Others are false positives. The env-example drift checker contains a regular expression for Ruby's ENV.fetch(. The npm lifecycle-script checker lists "fetch(" and "eval(" as strings to look for in other people's packages. A false refusal costs one tool. A false pass would let a calculator make network calls on an agent's behalf, so the trade-off is reasonable.
The runtime limit applies in MCP mode. Each call_tool and run_cli runs in a child Node process with a 20-second timeout and a 200 KB output cap, so a runaway loop or a huge result cannot hang or flood the client.
Be clear about what this is not. It is not an operating-system sandbox. The child process is ordinary Node, running code from your own clone. The gate keeps calculator engines from reaching the network or spawning processes. It does not make untrusted code safe. If you trust the clone, the gate keeps the tools to arithmetic.
Relation to the npm package
platform/distribution/mcp.mjs builds an npm package, monorepo-projects-mcp, that bundles the tools with a cli.mjs for npx users. Publishing it is off until the owner enables it, because publishing under the portfolio's name is an owner decision here. platform/agents/ needs no publishing: it works from any clone and covers every tool with an engine module, not only the ones with a CLI.
What is not known yet
Nobody outside the repository has used this yet as far as anything records. The repository has no measured visitors on any product (see Counting visitors without tracking them). Whether agents find tools by searching in a user's words is untested. If you try it, an issue saying which tool you called and whether describe told you enough would be the most useful thing to hear.