Skip to main content
Open Workflows, expand a workflow, then choose View production tracking. Tracking is available independently of simulation readiness. It reports observed starts and backend outcomes; it does not infer success from a successful tool call. Admins use Configure tracking to choose a workflow definition and operation namespace, name the start and backend success events, and set observation and correction windows. Optional failure and progress events explain where attempts stop. In Create operation namespace, grant each authenticated backend source only the permissions it needs. Browser sources cannot report authoritative outcomes. Use Authorize producers to add a source or permission to an already configured workflow’s namespace. Existing grants stay in place. After revoking a source in Sources, authorize its replacement here so current attempts keep their identity. Instrument each new business attempt with a new operation-instance reference after your application authorizes the operation. Participating servers must use the same namespace and workspace hash key. Emit the configured backend success event only after your application confirms the actual result.

Instrument your servers

Report attempts from your servers with the Go or Python SDK, on a milestone. Give each new business attempt a new reference, and the same reference for its start, its progress, its plugin calls and its outcome, across every service that takes part:
Send the role terminal with your configured success event only after your backend has confirmed the result. A tool call that returned, or an HTTP 200, is not completion, and the SDKs never assume it. A plugin call with no authorized operation sends only the environment and the plugin: it shows as plugin activity without workflow credit. The reference never leaves your server. The SDKs send a keyed hash of it, made with your workspace’s hash key, and both make the same hash, so services in different languages can report one attempt. Never put a raw reference in a milestone name, an invocation id or a log field. Evidence that cannot be accepted is left off that milestone and reported to OnError or on_error as invalid_performance; the milestone and your code still run. Completed within window includes only starts whose observation window has ended. Unknown outcomes remain in that group. Recent attempts are pending; an unconfigured workflow, no observed attempts, and delayed processing are shown separately. While new evidence is processing, displayed rates reflect the last completed evaluation. Rates are unavailable when attempts are overdue for evaluation or an operation exceeds processing capacity; the page explains the state. An operation over capacity does not hold up other operations; its retained evidence remains available, and new evidence triggers another evaluation. Unrelated workflows and test traffic do not hide a production rate. Production and test views stay separate, and each view uses one set of rules. In Plugins, register a ChatGPT or Claude plugin and connect its authenticated server deployment. Leave package fields empty when the installed package is unknown. A server deployment alone cannot establish which package a host installed. The plugin page shows participating workflow attempts and invocations without an eligible assignment. Participation does not establish causal impact.

Submission check

A plugin’s page has five tabs: Submission check, Submission pack, Usage, Deployments and Production tracking. The address keeps the tab you are on, so you can share a link to it. On a plugin’s page, Submission check tests the plugin’s MCP server against the directory it is for: ChatGPT’s plugin submission rules for a ChatGPT plugin, Claude’s connector directory rules for a Claude one. Anyone in the workspace can choose Run check and read the result. One check runs at a time for a plugin. It waits until a runner takes it, up to 30 minutes, and stops if it has not finished five minutes after that. The check signs in as a directory reviewer would, with the test account saved for the server on the Access tab in Setup: through your server’s own sign-in, or with a saved API key. Use an account that signs in with a password alone; a second factor stops the check. The page names the MCP server it uses: the Brand Vault MCP server whose address is your latest deployment’s server URL, or the one an Admin chooses under Brand Vault MCP server. With no test account saved, the check reads only what your server shows without signing in, and links to Access. The check shows how many rules pass and what the directory would refuse. Results are grouped by what they are about: Server, Tools, Sign-in, and Domain and public pages. A group with a failed rule or a warning opens by itself. Each result says whether it passed, failed, is a warning or couldn’t be checked, names the directories that have the rule, links to the page that states it under Rule, and says how to fix it. A rule a directory requires fails; one it only recommends is a warning. Copy fix prompt copies every failed rule and warning, with what the check saw and how to fix it, as one prompt to paste into your coding agent. Previous checks lists your earlier checks, newest first; choose one to read it. The check reads every tool, resource and prompt your server lists, your sign-in metadata, and your website, support, privacy and terms pages as Brand Vault names them, and, for ChatGPT, the domain verification file on your server’s host. It does not keep the token it signs in with. When a check cannot finish, the page says why. A plugin whose MCP server runs on your own computer or at a private address cannot be checked. The requests the check makes itself, to your server, its sign-in endpoints and your pages, stop at a private or internal address, including through a redirect. The browser that signs in with your test account follows your sign-in pages as any browser would. Your server’s address must start with https://. Tool hints differ by directory: Claude asks every tool for readOnlyHint or destructiveHint, and ChatGPT asks for readOnlyHint, destructiveHint and openWorldHint, each set to true or false. A tools list the check could not read whole is not judged, and its tool rules say they couldn’t check: your server answered a page with an error, or listed something the MCP specification does not allow, such as a tool whose input schema is not an object. Choose Inspect evidence to follow an attempt through its current journeys. The timeline retains events without a journey and labels evidence excluded from the outcome. A journey page also lists its related workflow attempts. Grouping changes update these links without changing the attempt’s business outcome.

Submission pack

Submission pack holds what you submit to the plugin’s directory. Choose Draft with AI to have one written from your Brand Vault and the tools the latest check of the plugin’s MCP server listed, or Write by hand to start empty. Read every field before you choose Save as new version: a draft can be wrong. Each field shows its directory’s limit, and the page refuses to save one past it. When a field to fix is on another step, saving takes you to that step. A pack goes step by step in the order the directory asks, each field with Copy, beside a preview of how the directory shows your listing. Its last step lists what the version still lacks before the directory takes it, including any rule the latest check failed, and then the directory’s own steps. A saved version never changes. To change one, choose Edit as a new version, which saves the next version and keeps the old one.
  • ChatGPT. A version is a plugin package: the listing, its links, five positive and three negative review cases, release notes and a demo video address. Choose Download ZIP and upload it in the OpenAI Platform’s Plugins, choosing your verified developer identity. The package name uses lowercase letters, numbers and single hyphens, and stays the same across updates. Every upload needs a version greater than any you have submitted, with no suffix such as -beta. A draft suggests one: a major version when a tool was removed since your last submitted version, a minor one when a tool was added, and a patch otherwise. ChatGPT reviews an MCP plugin only with all four links, the review cases, release notes and a demo video. Add the icon and the reviewer’s test account in ChatGPT’s dashboard: they are not in the package.
  • Claude. A version follows the steps of Claude’s developer portal: Connection, Listing (including the URL slug, which is permanent once published), Use cases (at least three), Company (including the contact Anthropic emails about the review), Authentication and Data handling. Copy each field into the portal and upload the icon there. The portal no longer takes reviewer instructions or a test account: if your listing goes to Verified review, the reviewer asks for test access by email, and Notes for a Verified reviewer keeps what you will send.
Set each version’s Status to what you did with it: Submitted, Live or Rejected. A version you have submitted cannot go back to Draft: save a new version instead. Making a version live marks the one live before it Replaced. Choose a version under Version to read it; when there are more than the list shows, choose Load older versions. Every day, a plugin with a live version gets a submission check of its own. When the plugin’s MCP server is not the one your live version was written for, or the tools it lists no longer match, because a tool was added or removed or its title, description, hints or input schema changed, the pack shows Changes since live with each change and Draft a new version. A move to another MCP server is a change even when the new server lists the same tools. A daily check that finds changes emails your workspace’s Admins, once for each set of changes to a live version.

Usage by client

A plugin’s page in Plugins has a Usage section: the MCP calls your plugin’s connected deployments answered in the last 30 days, per day and by the client that made them (Claude, ChatGPT, Codex, Cursor or Other). Claude Code counts as Claude, and a client that does not name itself counts as Other. It is built from the MCP calls your servers already record with the Node, Go or Python SDK, so there is nothing more to send. The section charts MCP calls per day and, once there are signed-in accounts, the accounts first seen per day. Counts update every hour, and the section says when they were last counted. Days are UTC days. Signed-in accounts counts each verified user once per client: the user your server named for a request, or the user an agent acted for. An app or agent acting for nobody is not an account. First seen counts the accounts first seen with that client in these 30 days; it is not a count of installs. Counting starts from the 30 days before a plugin’s first count, so in its first 30 days every account shows as first seen. Accounts appear only once your server names who is calling (Node, Go, Python); until then the section shows calls only. Calls by tool lists each tool, resource and prompt your server registered, by its calls in these 30 days. Failed results are tool calls that answered with a result marked as an error (isError); Failed calls are calls whose handler threw or returned an error. The Error rate counts both, among the calls whose outcome is known, and shows a dash when none is. A tool is named when your server records it through the SDK’s MCP helpers (Node 0.2.0, Go v0.2.0, Python 0.1.0 and later); calls that do not say what they called show as Unknown tool. The table lists the most-called first; choose Show more tools for the rest. A name longer than 256 characters is shortened and ends with its length.