ChatGPT Plexus Plugin

Connect ChatGPT to a self-hosted Plexus gateway. This repository includes two independently installable Agent Skills: plexus-admin and plexus-connect, in skills/<name>/.

Install the skills

Install either skill (or both) using the Vercel Skills CLI. For Hermes, select hermes-agent explicitly:

npx skills add gwyntel/chatgpt-plexus-plugin --skill plexus-admin --agent hermes-agent -g -y
npx skills add gwyntel/chatgpt-plexus-plugin --skill plexus-connect --agent hermes-agent -g -y

Omit -g to install into the current project instead of globally. Install both by running both commands. Other supported agents can be selected with --agent <agent>; list skills first with npx skills add gwyntel/chatgpt-plexus-plugin --list.

Update installed skills with npx skills update plexus-admin or npx skills update plexus-connect (add -g to update global installations). Review the resulting skill instructions after updating. Remove with npx skills remove plexus-admin / plexus-connect and the matching -g scope.

Related MCP server: Windows Console MCP

Use

Choose the installed skill for the task: Plexus management and administration (plexus-admin), or connecting to one configured remote MCP server through Plexus (plexus-connect). Skills provide operational guidance; they do not install or authenticate the ChatGPT plugins themselves. Follow the host application's connection flow and never provide credentials in chat.

Before using plexus-admin, an operator must replace the example setup URL in skills/plexus-admin/SKILL.md with the URL of their deployed private Site. Do not publish a Site or expose its administrator endpoint without reviewing its identity and network security. plexus-connect targets one configured MCP endpoint; it does not dynamically mount every upstream.

Management connection

The hosted adapter translates authenticated tool calls to Plexus management MCP requests at /mcp/plexus, supplying the administrator key server-side in x-admin-key. It does not add OAuth support to Plexus itself; the built-in Plexus OAuth service for proxied MCPs does not currently authorize its management endpoint.

Keys are stored in R2 encrypted with AES-256-GCM under a runtime secret, bound to the Site-scoped user identity. Browser writes require an authenticated, user-bound form token. Explicit foreign origins are rejected; null origins are supported for sandboxed or privacy-preserving browser contexts. Agent tools cannot set or retrieve the saved key.

The adapter has full administrator authority over the chosen instance. The initial Site is owner-private; publishing its source does not publish credentials or grant others access to the deployment. Do not make the deployed Site public without separately reviewing its authentication and network policy.

Source and hosting

• worker/index.js: responsive setup page, encrypted credential storage and stateless MCP endpoint.

• scripts/test-worker.mjs: local tests using a mock upstream and mock object storage.

• .openai/hosting.json: Sites deployment declaration. The project ID belongs to one deployment; remove it before registering your own Site.

• CREDENTIAL_ENCRYPTION_KEY: required 64-character hex secret, configured through hosted runtime settings. Keep it out of Git. Replacing it makes existing encrypted credentials unreadable.

• BUCKET: R2 binding used for user connection records.

This deployment uses Sites-managed identity headers. Do not run it behind an arbitrary proxy that accepts user-supplied oai-authenticated-user-id headers. A standalone deployment needs a verified identity boundary of its own.

Development and verification

node scripts/test-worker.mjs
bash scripts/build.sh
node scripts/validate-artifact.mjs

Publish through the Sites source workflow with the MCP capability enabled. Sites provisions the canonical plugin; do not create a second wrapper plugin for the same Site.

Remote MCP connections

For upstream servers such as Honcho, use Plexus's /mcp/<server> endpoint and its advertised OAuth flow. Upstream credentials belong in Plexus configuration, separately from the client-to-Plexus OAuth grant. The management adapter handles only the reserved /mcp/plexus endpoint.

Verification status

Local tests cover identity, null-origin form protection, encryption, isolation, edge-compatible redirect handling, protocol forwarding, response redaction and connection verification with mocked Plexus responses. On 2026-10-19, the owner completed browser setup and an authenticated read-only management status call succeeded through the installed plugin. Mobile layout uses responsive HTML; native mobile connection has not yet been separately user-verified.

Public repository export

The intended repository is gwyntel/chatgpt-plexus-plugin; the DeepWiki badge above points there. To prepare a clean public source tree without the private Site's project identity or Git history:

node scripts/export-source.mjs /absolute/new/output-directory

Inspect that tree before pushing. skills/plexus-admin holds the administration instructions and CLI. Its setup link is an example placeholder; replace it with your deployed private Site's URL before attaching the skill to your plugin. The source-pinned upstream CLI retains its MIT license in its scripts directory. Public GitHub source publication is separate from submission to the ChatGPT public plugin directory.

PUBLIC_ORIGIN must be set to the public HTTPS origin of your Site. It is used for redirects and origin checks behind the hosting proxy. The adapter does not follow upstream redirects, so configure the final HTTPS Plexus base URL.

Build a Plexus Connect package

For the second plugin, point a package at one upstream MCP exposed by Plexus:

node scripts/build-connect-plugin.mjs https://your-plexus.example/mcp/honcho /absolute/new/plexus-connect

The output contains portable manifests and the connection skill. Package that directory as one ZIP and import it through Plugin Creator. Configure OAuth on the Plexus instance first. This uses a fixed, explicit endpoint: template-only MCP URLs were not sufficient to register a usable connection. One package connects one named upstream; it does not dynamically mount every server behind Plexus.

Troubleshooting skill installation

• No skills found: use the repository slug exactly and --list; the supported discovery paths here are skills/plexus-admin/SKILL.md and skills/plexus-connect/SKILL.md.

• Invalid skill/frontmatter: use a current npx skills CLI and ensure each SKILL.md retains YAML name and description fields.

• Skill installed for the wrong agent or scope: repeat skills add with --agent hermes-agent and choose -g (global) or omit it (project). The install destination depends on both options.

• Permission errors: check write permissions for the selected agent's global skills directory or project directory; rerun with the intended user rather than using elevated privileges by default.

• Updates do not appear: update the same skill name and scope, then start a fresh agent session if it has already loaded an earlier copy.

License

Repository source remains under the existing MIT License; see LICENSE. The bundled upstream CLI retains its separate MIT license under skills/plexus-admin/scripts/LICENSE.