Toolars

Registre officiel Autres

Local-first dev utilities: hash, base64, JWT, UUID, cron, URL codec. No network or telemetry.

Installation

Ajoutez ceci à la configuration MCP de votre client (Claude Desktop : claude_desktop_config.json ; Cursor : .cursor/mcp.json ; VS Code et les autres utilisent la même forme « mcpServers »).

NPM package @toolars/cli
Claude Code
claude mcp add toolars -- npx -y @toolars/cli@0.1.1 -y -p @toolars/cli toolars-mcp
Configuration JSON (Claude Desktop, Cursor, VS Code)
{
    "mcpServers": {
        "toolars": {
            "command": "npx",
            "args": [
                "-y",
                "@toolars/cli@0.1.1",
                "-y",
                "-p",
                "@toolars/cli",
                "toolars-mcp"
            ]
        }
    }
}

Documentation

Tirée du README du projet sur GitHub, actualisée chaque semaine. Les liens ouvrent GitHub.

Sommaire
  1. Toolars open-source distribution
  2. Continuous verification
  3. What is published
  4. What is deliberately not published
  5. Verifying the local-processing claim yourself
  6. Running the CLI and MCP server
  7. Reporting a problem with this code
  8. License

Toolars open-source distribution

Use Toolars online for PDF, image, text, and developer tools directly in your browser.

This directory is the source of truth for the code Toolars publishes publicly. It is assembled into a distributable tree by scripts/export-open-source.mjs in the private repository; nothing here is built or bundled into the website.

Continuous verification

Every push and every day at 03:17 UTC, the Privacy proof workflow rebuilds both packages from this tree alone, re-runs their test vectors, and re-captures the privacy-proof receipt against the live site. The badge is green only when the latest receipt's outcome is verified.

A failed run is not automatically a privacy finding. The harness distinguishes two outcomes: violation means a measured request crossed the local-processing boundary (the workflow fails on the first attempt, without retrying); inconclusive means the run could not be completed — usually a network problem between the runner and the site — and says nothing about the site in either direction. Inconclusive runs are retried; a persistent one fails the job with an explicit notice. Read the run log before reading anything into a red badge.

What is published

PathWhat it is
evidence-harness/A runnable harness that drives the public website in headless Chromium and records every network request, including dedicated Web Worker fetches. It is the tool behind the claim that files never leave the browser.
packages/cli/@toolars/cli — a local stdio CLI and MCP server (hash, base64, JWT decode, UUID/ULID, timestamp, cron, URL encode/decode), shipped as a prebuilt bundle in dist/.
packages/local-tools/@toolars/local-tools — a smaller local stdio CLI and MCP server (JSON formatting, SHA-256), shipped as runnable source.

All of it is MIT-licensed (see LICENSE). THIRD-PARTY.md states the exact scope and lists third-party components.

What is deliberately not published

This is an open-core distribution, not the whole product. Not included:

  • the website application (routing, SEO layer, catalog, editorial content, the 13 localized content sets);
  • the tool implementations and Web Workers that the website ships, in source form — with one deliberate exception: packages/*/vendor/ carries the site runner cores these two packages import (15 files for the CLI, 2 for local-tools), because without them the packages could not be rebuilt or re-tested here. The packages/cli bundle also inlines compiled copies of the same cores, since it is the artifact prepared for a future @toolars/cli npm release. Both are covered by this repository's MIT license;
  • runtime assets (WASM engines, OCR language packs, fonts, PDF.js);
  • internal engineering, product, and audit documentation;
  • any deployment configuration, credentials, or analytics configuration.

That boundary is the point: the harness verifies the deployed site, so independent verification does not require the site's source. See evidence-harness/README.md for exactly what the harness does and does not prove.

Verifying the local-processing claim yourself

cd evidence-harness
npm install
npx playwright install chromium
npm run verify -- --base-url https://toolars.com

The run writes privacy-proof-receipt.json and exits non-zero if an unexpected request leaves the site's origin, if any request carries the canary marker, or if any request uploads data. Disclosed shell requests are counted separately. A missing or changing release identity also prevents a verified result.

Both packages are published on npm under the @toolars scope: npm install -g @toolars/cli gives you toolars + toolars-mcp, and npm install -g @toolars/local-tools gives you toolars-local + toolars-local-mcp. The bins do not collide.

Running the CLI and MCP server

cd packages/cli
node dist/cli.mjs hash --sha256 "hello"
node dist/mcp.mjs   # stdio MCP server

dist/ is committed and copied from the private build output, so running it needs no install and no build. Review that bundle if you want to see the code you would actually execute: it is unminified, and its section comments name the source file each part came from.

You can also rebuild it here and re-run the same test vectors the private workspace runs. Each package carries the exact site core closure it depends on under vendor/, so neither needs the private repository:

cd packages/cli
npm install          # esbuild + vitest, declared as devDependencies
npm run build        # writes dist/ from src/ plus vendor/
npm test             # shared digest vectors and CLI/MCP boundary tests
cd packages/local-tools
npm install          # esbuild + prettier
node build.mjs       # regenerates src/json-core.mjs from vendor/
node --test test/*.test.mjs

vendor/ holds copies of the site runner cores these packages import, produced by the exporter and byte-identical to the private ones. They are published as source on purpose: a reader who can rebuild the package and re-run its vectors can check the results instead of trusting them. Rebuilding in the exported layout changes source-path comments in the CLI bundles and local-tools generated core. Compare executable content separately from these comments; a rebuild is not claimed to reproduce the original manifest byte for byte.

Both packages are local-stdio only. Their inputs and results enter the calling client's context and may reach that client's model provider; the website's privacy promise does not extend to the client. Each package's README states this scope.

Reporting a problem with this code

Open an issue on the public repository. Security reports follow the process in that repository's SECURITY.md.

License

MIT — see LICENSE. Scope and third-party components: THIRD-PARTY.md and evidence-harness/fixtures/PROVENANCE.md.

Questions fréquentes

Qu'est-ce que Toolars ?

Local-first dev utilities: hash, base64, JWT, UUID, cron, URL codec. No network or telemetry.

Comment installer Toolars ?

Il s'exécute sur votre machine. Dans Claude Code : claude mcp add toolars -- npx -y @toolars/cli@0.1.1 -y -p @toolars/cli toolars-mcp. Pour les autres clients, collez la configuration JSON de cette page dans leurs réglages MCP.

Faut-il une clé API pour Toolars ?

Son entrée au registre n'indique aucune clé ni secret obligatoire. Certaines fonctions peuvent tout de même demander un compte auprès du service auquel il se connecte.

Toolars est-il open source ?

Oui. Son code est sur GitHub sous licence MIT.

Qui maintient Toolars ?

Il est publié par aixtral. AI.tn le référence sans le maintenir ni le recommander.

Issu du registre MCP officiel. Être listé n'est pas une recommandation : vérifiez le code et les permissions d'un serveur avant de lui donner accès à vos données.