Marketplace Management

Marketplace management lets platform administrators and maintainers control which assistants appear in the Marketplace. Assistants submitted by users are validated automatically and then wait in a review queue until an administrator approves or rejects them. Administrators can also audit the quality and usage of published assistants and configure the validation thresholds.

For the submitter's side of the flow, see Publish to Marketplace.

Enabling Marketplace Management

Marketplace management is off by default and is turned on per environment with the MARKETPLACE_MANAGEMENT_ENABLED and MARKETPLACE_APPROVAL_GATE_ENABLED environment variables of the codemie-api service. Parameters and defaults are listed in Marketplace Management parameters.

Reviewers are users with the administrator or maintainer role. Without at least one such user, every submission stays in PENDING_REVIEW with no way to resolve it.

Accessing Marketplace Management

Marketplace management is available to users with the platform administrator or maintainer role.

1. Click the Profile icon in the bottom left corner and select Settings.
2. In the Administration section, select Marketplace management.

The page contains the following tabs:

Tab	Purpose
Overview	Marketplace quality audit results and statistics
Review queue	Assistants waiting for an administrator's decision
Cleanup queue	Published assistants sorted by quality and usage, with recommendations for removal
Config	Usage thresholds, validation gates, and similarity settings used by the checks and the audit

Overview

The Overview tab shows the results of the latest Marketplace audit. The audit sorts every assistant published to the Marketplace into a bucket based on its usage and quality-gate results. The results are shared across all administrators; the time of the last audit is shown at the top of the tab.

Running the Audit

The audit runs only when it is started manually. Run it after new assistants are published or edited, and after changing the settings on the Config tab, so that existing assistants are evaluated again.

1. On the Overview tab, click Run audit.
2. Wait for the audit to finish. The audit runs in the background and can take several minutes on a large Marketplace. The page can be left in the meantime; the numbers update once the audit finishes, and all administrators see the same run.

The audit only assigns buckets and recommendations. It never removes assistants from the Marketplace; removal is always a manual action in the Cleanup Queue.

Buckets

The three cards at the top show how many assistants are in each bucket:

Card	Assistants in the bucket
Qualified	Passed the core quality gates and are in the Hot or Warm usage tier (Warm with at most one polish gap), or in the Cold tier when fully polished and used by at least 2 users.
Community	Remain in the Marketplace but are flagged: gate failures, minor polish gaps, or an assistant within its grace period.
Flagged for removal	Recommended for removal: a cleanup verdict (duplicate, invalid, or unused), part of an oversized cluster of similar assistants, a Cold-tier assistant that failed core gates, or a Cold-tier assistant that stalled past its grace period without growing.

Each card also shows the share of all Marketplace usage that falls on the assistants in the bucket, calculated from the number of unique users of each assistant. If no usage data is available, the card shows Usage data not available. Click a card to open the Cleanup queue filtered to that bucket.

Usage tiers, polish gaps, and grace periods are defined by the settings described in Configuration.

Statistics

Counter	Description
Total Marketplace Assistants	Number of assistants in the Marketplace.
Pending review	Number of assistants waiting in the Review queue.
Failing a critical check	Number of assistants with at least one CRITICAL finding.
Failing a minor check	Number of assistants with at least one OPTIONAL finding, such as a description that repeats the name.

Flagged Assistants and Failed Checks

The tables below the counters break down the audit results:

• Assistants flagged, by reason: the number of assistants flagged for each reason, for example:
  • Fails quality checks: the assistant has two or more CRITICAL findings and is in the Cold or Dead tier.
  • Unused: the assistant is in the Dead tier and has not been updated for longer than the Dead grace period.
  • Duplicate: the assistant's similarity score with another assistant reaches the Similarity Critical threshold.
  • Duplicate purpose: the assistant's purpose is very similar to that of a more widely used assistant.
• Failed quality checks: each failed check (for example, Instructions too short, Description just repeats the name, or Possible credential in the instructions) with its importance (Critical or Minor) and the number of assistants that fail it.
• Most-used MCP servers and Most-used categories: the MCP servers and categories used by the most Marketplace assistants, with the number of assistants and their total users.

Click Filter → in any row to open the Cleanup queue with the corresponding filter applied.

Review Queue

The Review queue tab lists assistants that passed the automated validation checks and are waiting for an administrator's decision. The number next to the tab name shows how many assistants are pending review. Assistants in the queue have the REVIEW status and are not visible in the Marketplace. Gate findings are recalculated every time the list is opened, so a pending assistant can show CRITICAL findings, for example after the thresholds on the Config tab were raised.

Finding Assistants in the Queue

Use the controls above the table to narrow down the list:

Control	Description
Search	Search assistants by name or creator.
Category	Show assistants assigned to a specific category.
Wait time	Show assistants by how long they have been waiting: All Wait Times, < 1 day, 1-7 days, 7-30 days, or > 30 days.
Sort	Change the order of assistants in the table: Default order, Submitted (newest first), Submitted (oldest first), or Name (A-Z).

The table shows the following columns:

Column	Description
Assistant	Assistant name.
Submitted by	User who submitted the assistant for review.
Submitted at	Time elapsed since the assistant was submitted.
Category	Categories assigned to the assistant.
Actions	Menu (⋮) with the review actions.

Reviewing Assistant Details

Click an assistant name in the table to open the Assistant details dialog:

The dialog contains the following information:

• Submission: the assistant name, description, and categories.
• Submission metadata: Status, Visible in Marketplace, Submitted by, Submitted at, Creator, Project, Created, Updated, Decided by, and Decided at.
• After a decision: Decision reason and Bypassed gate findings, which show who overrode the gate and which CRITICAL results were bypassed.
• Quality gates: the number of Critical and Minor gate failures found for the assistant.
• Capability: the toolkits and other capabilities of the assistant. Click Show system prompt to view the system prompt.
• AI Assessment: click Assess with AI to score the assistant across five quality dimensions and get concrete suggestions for improvement:
  • Name, description and category
  • Description quality
  • Instruction quality
  • Category fit
  • Uniqueness

Use the buttons at the bottom of the dialog to make a decision: Publish and Reject, or Force-publish and Reject for an assistant with CRITICAL findings. They work the same way as the row actions described below.

Approving an Assistant

1. In the Actions column, click ⋮ next to the assistant.
2. Select Publish.

The assistant is published in the Marketplace with the VERIFIED status and is removed from the review queue.

Force-Publishing from the Review Queue

For a pending assistant with CRITICAL findings, the row menu and the details dialog offer Force-publish and Reject instead of Publish.

1. In the Actions column, click ⋮ next to the assistant and select Force-publish.
2. In the Force-publish assistant popup, enter the justification in the Override notice field. The notice must be 20–2000 characters long.
3. Confirm the action.

The assistant is published with the VERIFIED status. The notice and the bypassed CRITICAL findings are stored with the decision and shown as Decision reason and Bypassed gate findings in the details dialog.

Rejecting an Assistant

1. In the Actions column, click ⋮ next to the assistant.
2. Select Reject.
3. In the Reject assistant dialog, enter the reason in the Reason field. The reason must be 20–2000 characters long; the dialog shows the current character count.
4. Click Reject.

The assistant gets the REJECTED status and is removed from the review queue. The submitter sees the reason by hovering over the REJECTED badge on the assistant card, as described in Publish to Marketplace.

tip

Write the rejection reason for the submitter: state which requirement is not met and what to change, so the assistant can be fixed and submitted again.

Reviewing Several Assistants at Once

1. Select the checkboxes next to the assistants. To select all assistants in the table, select the checkbox in the table header.
2. Above the table, the number of selected assistants and the bulk actions appear:
   • Approve selected: publish all selected assistants.
   • Reject selected: reject all selected assistants with one shared reason.
3. To clear the selection, click × next to the number of selected assistants.

To reject several assistants:

1. Select the assistants and click Reject selected.
2. In the Reject N assistant(s) dialog, enter the reason in the Reason field. The reason must be 20–2000 characters long and is shared with all submitters of the selected assistants.
3. Click Reject.

Cleanup Queue

The Cleanup queue tab lists the assistants published to the Marketplace together with the latest audit results: usage tier, bucket, and decision. Use it to find assistants that are unused, duplicated, or of low quality and to unpublish them. The number next to the tab name shows how many assistants are in the queue.

Filtering and Sorting

Changing a filter applies it immediately. When any filter is active, Clear all appears to reset all filters. Filters opened from the Overview tab (for example, a failed check) appear as removable chips above the table.

Filter	Options
Search	Search assistants by text.
Bucket	All (default), Qualified, Community, or Delete.
Reason	All Reasons, or one of the reasons that currently have assistants: Duplicate, Duplicate purpose, Near-identical copy, Fails quality checks, or Unused.
Tier	All Tiers, Hot, Warm, Cold, or Dead.
Decision	All Decisions, Keep, Review, or Cleanup.
MCP	All MCPs, or an MCP server name with the number of assistants that use it. Shows only assistants that use the selected MCP server.
Category	All Categories, or a category with the number of assistants in it.

To sort the table, click a column header:

• Assistant: A→Z, then Z→A, then unsorted.
• Users and Likes: highest first, then lowest first, then unsorted.

Understanding the Columns

Column	Description
Assistant	Assistant name and categories. For a duplicate, the line ↳ canonical: shows the original assistant it duplicates.
Creator	User who created the assistant.
Users	Number of unique users of the assistant.
Likes	Number of likes.
Tier	Usage tier based on the number of unique users.
Bucket	Marketplace qualification bucket assigned by the audit.
Decision	The audit's recommendation for the assistant.

Hover over the ⓘ icon next to a column name, or over a tier or bucket chip in a row, to see its description.

Tier is based on the number of unique users of the assistant. With the default Marketplace Thresholds:

Tier	Unique users
Hot	50 or more
Warm	7–49
Cold	3–6
Dead	fewer than 3

Bucket has the same meaning as on the Overview tab: Qualified, Community, or Delete (shown as Flagged for removal on the Overview tab).

Decision shows the audit's verdict for the assistant:

Decision	Meaning
Keep	The assistant itself needs no changes. Check Bucket for its Marketplace visibility status.
Review	The assistant requires human review.
Cleanup	The assistant is recommended for removal. Hover over the ⓘ icon next to the decision to see the details.

The Reason filter values explain why an assistant is recommended for removal:

• Duplicate: the similarity score with another assistant reaches the Similarity Critical threshold.
• Duplicate purpose: the assistant's purpose is very similar to that of a more widely used assistant.
• Near-identical copy: the assistant is a near-identical copy within a group of similar assistants.
• Fails quality checks: the assistant has two or more CRITICAL findings and is in the Cold or Dead tier.
• Unused: the assistant is in the Dead tier and has not been updated for longer than the Dead grace period.

Viewing Assistant Details and AI Assessment

To open the Assistant details dialog, click the assistant name, or click ⋮ and select View details. The dialog is the same as in the Review Queue, with an additional Decision section and an Unpublish button.

To score the assistant's quality, click Assess with AI in the AI Assessment panel, or click ⋮ and select Assess with AI to open the dialog and start the assessment right away. The assessment shows:

• Overall: the overall score and the recommended bucket (for example, Community).
• Quality scores: a score from 0 to 100 with an explanation for each of the five dimensions: Name, description and category, Description quality, Instruction quality, Category fit, and Uniqueness.

Once a result exists, the button changes to Re-assess.

Unpublishing Assistants

Unpublishing removes an assistant from the Marketplace listing and the search index and returns it to its creator. The assistant is not deleted.

To unpublish one assistant:

1. Click ⋮ next to the assistant and select Unpublish, or click Unpublish in the Assistant details dialog.
2. Confirm the action.

To unpublish several assistants:

1. Select the checkboxes next to the assistants. The checkbox in the table header selects all assistants on the current page only.
2. Click Unpublish selected.
3. Confirm the action in the Unpublish N assistant(s)? dialog.

Configuration

The Config tab contains the thresholds used by the validation checks and the Marketplace audit. The settings are grouped into sections; each section has its own Save button.

The label in the top right corner of each section shows where the values come from:

• Default from config: the section uses the platform default values listed below.
• Overridden: the values were changed and saved on this tab. Click Reset to default next to Save to return to the default values.

Changes take effect on the next audit run. Existing assistants are not re-evaluated automatically: after changing the settings, click Run audit on the Overview tab.

Marketplace Thresholds

Usage tiers and grace periods used by the audit and the Cleanup queue. Tiers are based on the number of unique users of an assistant: Hot from the Hot tier minimum, Warm from the Warm tier minimum up to the Hot tier minimum, Cold from the Cold tier minimum up to the Warm tier minimum, and Dead below the Cold tier minimum.

Setting	Default	Description
Hot tier min. users	50	Minimum active users for the Hot tier.
Warm tier min. users	7	Minimum active users for the Warm tier.
Cold tier min. users	3	Minimum active users for the Cold tier.
Cold grace period (days)	60	Days in the Cold tier before cleanup.
Dead grace period (days)	30	Days in the Dead tier before cleanup.

Core Gates

Minimum requirements checked when an assistant is published. Failing a core gate is a CRITICAL finding.

Setting	Default	Description
Description min. length	100	Minimum number of characters required in the assistant's description.
System prompt min. length	200	Minimum number of characters required in the system prompt.
Substance-prompt alt min. length	500	Minimum number of characters required in the system prompt of an assistant without tools, MCP servers, skills, or data sources.

Polish Gates

Additional quality requirements checked only during the audit. Polish gates are not checked when an assistant is published and do not block approval; they affect which bucket a published assistant is placed in (for example, Qualified or Community).

Setting	Default	Description
Description min. length	250	Minimum number of characters required in the assistant's description for the polish check.
Categories min. count	2	Minimum number of categories the assistant must be assigned to.
Require icon	Enabled	Whether the assistant must have an icon to pass the polish check.
Conversation starters min. count	2	Minimum number of conversation starters the assistant must have.

Similarity

Settings for comparing assistants with each other. Similarity scores range from 0 to 1.

Setting	Default	Description
Critical threshold	0.70	Similarity score above which two assistants are treated as near-duplicates.
Optional threshold	0.55	Similarity score above which two assistants are flagged as worth reviewing for overlap.
Report threshold	0.40	Similarity score above which a pair appears in similarity reports.
Shortlist size	10	Maximum number of candidate assistants to compare against each assistant during a similarity check.

Consolidation

Settings for clusters of similar assistants. When a cluster of similar assistants grows beyond the minimum size, only the most widely used ones are kept and the rest are flagged for removal in the audit.

Setting	Default	Description
Big cluster min. size	5	Minimum number of similar assistants for a group to be treated as a big cluster.
Keep top N per cluster	3	Number of assistants kept in a big cluster; the others are flagged for removal.