ToolsEnabled Bench

Build the benchmark.
Keep the evidence.

Compose prompts from reusable parts, freeze the protocol, and export a study whose reports can be regenerated from retained evidence. A local research workbench with a browser dashboard and twelve tools for MCP clients. No ToolsEnabled account is required.

Download Bench 0.3.3 beta Read the source

ToolsEnabled Bench 0.3.3 beta · MIT licensed · Node.js 22.19 or later

One benchmark family. Engine Bench and Head Gasket Bench keep their own tasks and scoring. This page is the workbench for the next study.

Study workflow: reusable prompt parts become a frozen protocol, with retained responses, logs and reports

How it works

Compose. Freeze. Inspect.

Build a study

Combine typed snippets and nested structures. Declare choices, seeds, exclusions and the analysis plan. Start with the recorded arithmetic example or the structured-record example; neither requires a model account.

Freeze the protocol

Save an identity for the specification, runtime, required plugins, prompts and schedule. Export them together with a standalone CLI and file hashes.

Keep the evidence

Inspect attempts, responses, logs and scoring evidence. Regenerate HTML and Markdown reports from the retained run. A new stochastic model response is not expected to match a previous one.

Local data

On your computer.
Under your control.

Projects, frozen studies, logs and reports stay in the private data folder you choose. The dashboard listens only on 127.0.0.1. The MCP server runs locally over stdio. Bench has no account service, telemetry or automatic upload.

Execution uses your permissions

A collector you explicitly run can send prompts and declared inputs to its configured endpoint or invoke local commands and plugins. Those programs run with your user account’s filesystem and network permissions. Execution confirmation and file hashes do not provide an OS sandbox.

Your client receives tool results

Your MCP client may send results to its model provider under that provider’s policies. Saved evidence and exported studies may contain private data; review them before sharing.

Keep the whole record

Back up the whole data folder and retain the matching trusted runtime for old studies. Read the release privacy notice for storage and retention details.

Install

Download. Verify. Start.

Version 0.3.3 beta was published on 15 October 2026. The same runtime ZIP serves the dashboard and local MCP server. It includes the prebuilt application and pinned MCP SDK. Install Node.js 22.19 or later separately; the runtime archive needs no npm install.

Expected runtime ZIP SHA-256: ece76defcbd817262faff6a8391f5b77fe17343359733e1129c21677e7afdce2. Compare before extraction. Obtain the archive from a source you trust; checksums establish integrity, not publisher authentication. Use a fresh installation folder and a private local data folder outside it. Stop if any step fails.

Linux and macOS — terminal

Download the ZIP and its sidecar into one fresh folder. Linux uses sha256sum; on macOS use the indicated shasum check instead. Continue only when the checksum reports OK.

node --version
# Linux
sha256sum -c toolsenabled-benchmark-builder-0.3.3.zip.sha256
# macOS: use this check instead
shasum -a 256 -c toolsenabled-benchmark-builder-0.3.3.zip.sha256
unzip toolsenabled-benchmark-builder-0.3.3.zip
cd toolsenabled-benchmark-builder-0.3.3
node tools/release.mjs --verify
export BENCHMARK_DATA_DIR="$HOME/ToolsEnabled-Bench-data"
node server/main.mjs

Windows — PowerShell

Compare the PowerShell hash with the sidecar before extraction. Then run the remaining commands in the same window.

$ErrorActionPreference = 'Stop'
node --version
Get-FileHash .\toolsenabled-benchmark-builder-0.3.3.zip -Algorithm SHA256
Get-Content .\toolsenabled-benchmark-builder-0.3.3.zip.sha256

Expand-Archive .\toolsenabled-benchmark-builder-0.3.3.zip -DestinationPath .\bench-install
Set-Location .\bench-install\toolsenabled-benchmark-builder-0.3.3
node tools/release.mjs --verify
if ($LASTEXITCODE -ne 0) { throw 'Bench archive verification failed.' }
$env:BENCHMARK_DATA_DIR = Join-Path $env:USERPROFILE 'ToolsEnabled-Bench-data'
node server/main.mjs

Run a recorded study

Open http://127.0.0.1:4318. Choose A small, complete study, then Freeze project and Run frozen study locally. The example uses recorded responses; inspect its reports in Local runs. Leave the terminal open; Ctrl+C stops the dashboard.

MCP hosts

Connect your agent.

Stop the dashboard first. From the extracted runtime, with BENCHMARK_DATA_DIR still set, print the configuration for your client. The helper prints setup; it does not change your client.

Claude Code

Prints a terminal registration command. Review it, run it in the matching shell, then open a new session.

node tools/mcp-config.mjs --client claude

Claude Desktop

Prints JSON for the client configuration. Merge it with existing entries. This is the Node/stdio route; native Desktop bundle acceptance remains unverified.

node tools/mcp-config.mjs --client claude-desktop

Codex

Prints a terminal registration command. Review it and run it in the matching shell. For long studies, set a suitable tool_timeout_sec in the Bench MCP configuration.

node tools/mcp-config.mjs --client codex

Cursor

Prints JSON for MCP configuration. Review the paths and merge it with existing entries.

node tools/mcp-config.mjs --client cursor

DeepSeek Harness

Prints a Cordis overlay entry. Merge it into the appropriate plugin list; keep your existing configuration.

node tools/mcp-config.mjs --client deepseek

The twelve tools cover saved projects, atoms, compositions, task generation, freezing, export, qualification, running, analysis and reports. Execution requires the exact study ID as confirmation; studies frozen elsewhere also need explicit trust after review. See the MCP reference and host setup guide. The previously documented marketplace is unavailable; use direct stdio registration.

Evidence and limits

Know what the evidence establishes.

Research apparatus and recorded controls

This beta supplies research apparatus and authored recorded controls. Hashes establish content identity. They do not establish independent scientific validity, prospective registration or authenticated provider identity.

Review before execution

Inspect a received study through an already trusted Bench installation. Running the received study’s own CLI executes its code before it can verify anything.

Optional apparatus has its own limits

Optional native domain apparatus has separately declared dependencies and limits; it is not required for the recorded examples. Native Lean/Docker campaigns and live-model collection remain unverified for this release.

Preserve the original study

Cite the software version and each study separately. Historical frozen records keep their original generator identities and matching runtime.

Open source.
MIT licensed.

Read the code, run the source tests and report issues on GitHub.

View Bench on GitHub Support and issues

Current release: ToolsEnabled Bench 0.3.3 beta · Release notes · Security policy · Privacy · MIT license

Released as tag v0.3.3 from commit 988e6705cfa66cbac7176f99b59b8e9760b2dd9d. The release records 585/585 exact-source tests locally and in GitHub CI, browser checks and independent Linux clean-folder review. Two clean source builds and an extracted-runtime repack produced identical archive bytes; all seven release assets were downloaded back and matched. See VERIFICATION.json for scope and toolchain.

Build the benchmark.Keep the evidence.