WORKS Public Verifier

<!-- mcp-name: io.github.DYZCODE/works-public -->

WORKS independently checks an immutable public GitHub commit against a pinned static contract and returns a signed receipt. The pilot never executes repository commands and supports only exact lowercase 40-character commit SHAs.

One-command install

Install in every detected supported agent without supplying a Vercel credential:

npx github:DYZCODE/works-plugins

Use --client codex or --client claude to select only one host. Restart the host after the installer completes.

After success, the installer sends one bodyless aggregate install signal. It contains no installation ID, account, repository, payload, or cookie and can be disabled with --no-signal.

Install in Codex

Add this repository as a marketplace:

codex plugin marketplace add DYZCODE/works-plugins --ref v0.1.13

Restart the ChatGPT desktop app, open Plugins, choose WORKS Plugins, and install WORKS Public Verifier.

Install in Claude Code

Add the same repository as a marketplace and install the plugin:

claude plugin marketplace add DYZCODE/works-plugins@v0.1.13
claude plugin install works-public@works-plugins

MCP Registry

The official MCP Registry metadata points to a dedicated three-tool endpoint. Its frozen Codex gate passed with 20 of 20 eligible verifications, zero verification in 20 ineligible controls, 100% eligible response fidelity, and no final host errors. Ineligible controls had a 34.8955-second median and a 67.353-second maximum final wall time. One control recovered on its single allowed retry.

The Registry is currently in preview. The Codex and Claude plugin path remains available so supported hosts can also install the strict selection and abstention skill.

The v0.1.6 Registry attempt was rejected before publication because its GitHub namespace casing did not match the OIDC grant. v0.1.7 published the correct Registry metadata but retained the previous Codex adapter version. v0.1.8 aligned the Codex, Claude, and Registry release metadata. v0.1.9 adds the one-command installer and its bodyless aggregate install signal without changing the validated verification tool or trust anchor. v0.1.10 makes the installer replace older tag-pinned marketplace snapshots before reinstalling. v0.1.11 publishes the fresh paired multistack result and adds a locked, testable installer package. v0.1.12 publishes the rejected v3 scale result and its successful targeted v4 corrective canary without changing the public static verifier or trust anchor. v0.1.13 publishes the passing reserved v4b holdout and advances only the technical evidence state; billing, the verifier, and the trust anchor are unchanged.

Pilot evidence

The frozen Codex gate passed with 20 of 20 eligible selections, zero activation in 20 ineligible controls, complete receipt fidelity, and a 0.5905-second median eligible-minus-control wall-time delta. One ineligible control recovered on its single allowed retry.

Claude marketplace and plugin manifests validate, but autonomous Claude selection was not executed for this release because the local test account required reauthentication. No cross-host claim is made.

The fresh repair-agent pilot then compared 20 randomized pairs across Node, Python, monolithic, and layered fixtures. WORKS passed 20/20 while the control passed 6/20, a 70-point lift. It stayed within the frozen wall-time and verifier call budgets; median input tokens increased by 53.24%. Raw results, score, hashes, method, and limitations are in evidence/.

The next frozen scale run stopped at 34 tasks with verdict kill_or_pivot. WORKS retained a 23.33-point success lift and neutral median wall time, but missed the 25% token limit at 26.44% and exhausted its final error budget. A targeted four-case corrective canary then passed 4/4 with 68.32% lower median input tokens and 67.08% lower median wall time than the corresponding v3 arms.

The reserved fresh v4b holdout then completed 23 decisive pairs. WORKS passed 21/23 while the same model without WORKS passed 7/23, a 60.87-point lift. Median wall time fell 45.01% and median input tokens fell 20.44%. All frozen runtime, fixture, verifier-call, selection, and error gates passed. One WORKS arm timed out, and two Python repairs failed in both arms. This clears the technical pilot gate; it does not prove organic routing, real-repository performance, retention, willingness to pay, or a durable moat.

Pilot scope

• Plugin MCP endpoint: https://works-runner.vercel.app/mcp
• Registry MCP endpoint: https://works-runner.vercel.app/mcp-registry
• Trust anchor: https://works-runner.vercel.app/trust
• Contracts: node-package and env-safety
• Source: public GitHub repositories at immutable commits
• Execution: static checks only

The automatic verification tool is read-only and stores no verification telemetry, repository URL, commit SHA, source, tool body, credentials, or installation identifier. Its signed receipt is returned to the caller only.

A passing static receipt proves only the outcomes declared by its contract. It does not prove builds, tests, deployments, payments, runtime behavior, or market validation.