Nymrel Plugin MCP Server

Nymrel universal plugin

Public workflow resources

The Nymrel Workflow Starter Kit contains two standalone skills—Evidence Workspace and Capability Adoption—and a Material Change Digest template. It includes offline code, synthetic examples, and an explanation of how we use each workflow. This source preview works independently of the hosted Nymrel connection; see its release notes for checks and publication boundaries.

Download the standalone source kit (SHA-256).

One installable Nymrel front door for ChatGPT and Codex, with a compatible Claude package and one hosted MCP service.

Tag Nymrel

After the plugin is installed, start a new conversation and tag @Nymrel:

• @Nymrel turn this idea into a build brief.
• @Nymrel audit https://example.com and prioritize the fixes.
• @Nymrel help me choose the right service for this business.

The nymrel-studio skill handles client discovery and creates a scoped brief without pretending it was sent. The nymrel-website-audit skill and hosted MCP tools handle the production read-only workflows. A future authenticated handoff can submit approved briefs and return client status through the same tag.

The operator connection already works in ChatGPT and Codex. Public availability still requires OpenAI directory review and publication; a local installation or developer-mode connection is not public distribution.

Live bindings

• ChatGPT app: the .app.json file binds the installed plugin to the registered Nymrel app.
• Portable plugin: plugin.json and mcp.json declare Agent Plugins 1.0 metadata and the existing hosted connection; both skills and the existing app binding are preserved.
• Codex plugin: the .codex-plugin/plugin.json compatibility manifest discovers both skills, the app, and remote MCP configuration.
• Claude plugin: the .claude-plugin/plugin.json manifest discovers the remote MCP configuration; Claude publication remains separate.
• MCP: every host connects to https://mcp.nymrel.com/mcp through the root .mcp.json file.

The hosted service is the source of truth for production tools. The TypeScript server in this folder remains a local development harness for the polished Website Audit widget; it is not the production MCP endpoint.

The hosted Python service now also contains a separately activated audit app, explicit public result schemas, and resumable-read guidance. Its view consumes the real hosted result shape; it does not depend on the TypeScript harness. See the implementation record.

The public tool workbench adds a shared view for all four tools, with global and thread entrypoints, browser display preferences, and explicit sharing of results into conversation context. It uses the existing APIs and makes no data-tool call on launch. NYMREL_TOOLS_UI_ENABLED is independently disabled by default pending host acceptance. See the extension record.

Product boundary

• Audits one public HTTP(S) page plus conventional discovery files on the same origin.
• Performs read-only technical checks. It does not log in, crawl private pages, change the target site, or claim legal/compliance certification.
• Blocks local, private, link-local, reserved, credential-bearing, and non-web targets before fetching.
• Returns useful structured data and text even when the host does not render the widget.

Local widget development

fnm use (Get-Content .node-version)
corepack npm@11.19.1 ci
corepack npm@11.19.1 run verify
corepack npm@11.19.1 audit --audit-level=high
corepack npm@11.19.1 audit --omit=dev --audit-level=high
corepack npm@11.19.1 start

The repository pins Node 24.20.0 and npm 11.19.1 while continuously proving compatibility on Node 22.22.0 and 24.20.0. Install scripts are denied unless an exact package version appears in allowScripts; the current reviewed set is only esbuild@0.28.2.

The HTTP MCP endpoint is http://127.0.0.1:8787/mcp. A standalone widget preview is available at http://127.0.0.1:8787/preview.

For stdio-only development hosts, build first and launch node dist/server.js --stdio directly. The packaged .mcp.json intentionally points at the live HTTPS endpoint.

Production tools

• nymrel_audit_website
• nymrel_find_domain
• nymrel_golf_bag_gap
• nymrel_social_clip_score

All four tools are live, credential-free, read-only, and backed by the production Nymrel API. Trade evaluation, permit lookup, and studio intake remain outside the public catalog until their provider or authorization contracts are production-ready.

The hosted source also includes an activation-held private-handoff adapter for approved clients. It keeps @Nymrel as the invocation surface while requiring OAuth and server-side project allowlists for submit/status calls. The production code path is pinned to an RS256 OAuth issuer and the exact MCP audience, but the feature remains disabled until a Nymrel-owned AuthKit tenant, tester subjects, dedicated storage, and production canaries satisfy the release gate. See hosted-mcp/README.md for the activation boundary.

The deployable source for https://mcp.nymrel.com/mcp lives in hosted-mcp/. The broader vendored API contracts stay private to that deployment package; hosted-mcp/api/index.py explicitly publishes only the four production-ready tools.

Validation levels

• Static contract: plugin manifest, exact tool metadata, MCP Apps resource metadata, and eight review evals.
• Compile/unit: npm run check and npm test.
• Runtime: npm run build && npm run test:mcp starts the real HTTP server, connects an MCP client, and verifies tool/resource discovery.
• Hosted runtime: Python 3.13 runs the pinned FastMCP service through compile, Ruff, packaging, OAuth, review-case, and dependency-audit gates on Linux and Windows.
• Live public fetch: npm run test:live audits https://example.com; use only when outbound network proof is wanted.
• Host: connect the deployed HTTPS /mcp endpoint in ChatGPT Developer Mode, invoke @Nymrel, and run the prompts in evals/review-cases.json (six positive, four negative, one per published tool).

Publication boundary

OpenAI now uses one universal plugin directory for supported ChatGPT and Codex surfaces. The Nymrel package has the required manifest, registered app binding, skills, production MCP URL, review cases, and publisher links. It is not public until the OpenAI submission portal records an approved, published version.

Refresh the installed ChatGPT tool catalog before submission and compare it with the deployed endpoint. The 2026-09-29 production probe found four public tools plus seven OAuth-protected Remote read tools; it found no UI resources. The new audit app remains disabled by default until target-host acceptance. Anthropic review remains separate.

See CONTRIBUTING.md for the complete local gate and SECURITY.md for private vulnerability reporting. GitHub Actions execution is separate evidence: a local pass does not imply that a hosted workflow, deployment, or directory review has completed.