Skip to main content
Glama
Plugin & Skill Manager

Plugin & Skill Manager

Plugin & Skill Manager

Release License GitHub stars

See which plugins and Skills are installed and used, identify redundancy, and streamline your tool environment for more efficient, controllable work.

  1. Inventory your tools in one place
    Inspect plugin and Skill sources, versions, installation and enablement status. Spot duplicate installations, outdated versions and stale entries to reduce lookup and maintenance effort.

  2. Track usage to inform decisions
    Use cross-session records, time-based statistics and traceable sources to understand which tools are used frequently and which have had no records for a long time.

  3. Identify redundancy and support safe cleanup
    Consider overlapping functionality, duplicate installations, usage and your needs to identify plugins or Skills that could be uninstalled. You decide what to remove, reducing tool-selection noise, maintenance effort and unnecessary permissions. When cleanup reduces service startup, tool definitions or loaded instructions, it can improve runtime efficiency and reduce token consumption.

A local activity dashboard for Codex: calendar heatmaps, date-filtered rankings, and separate records for Skill adoptions and plugin call attempts.

English | 中文 · Install · Latest release · Glama directory · Apache-2.0

v0.4.14 English activity dashboard showing 1,504 real recorded call attempts across 15 days and 8 tool services

What it helps you do

  • Compare recorded activity in the last 30 days and over a year. Bars and calendar squares share a selected date; pinned daily rankings list names and counts. Back to all restores the period filters.

  • Switch between Skill adoptions and plugin tool-call attempts without mixing the metrics.

  • Filter period rankings by date independently of the pinned day, and distinguish missing coverage from zero activity.

  • Inspect low-use candidates and items with no records in a compact observation panel. Open the counting details when needed; nothing is archived automatically.

  • Review structural redundancy candidates in Manage: duplicate content, same-name instances and plugin-version candidates. Keep or defer items without treating a candidate as proof that removal is safe.

  • Switch between Plugins and Skills in Manage. Plugin inventory counts unique IDs, supports installation-status filters and name/marketplace/version search, and distinguishes installed versions from cached Skill files. Installed plugins without Skills are included. Refresh rechecks installation status; failures retain the previous result and mark it stale.

  • Click Personal skills, System skills, Skills in plugin cache, or Total in Manage to browse that category, then search within it. Counts represent Skill files, not plugins.

  • Plugin installation status is checked automatically when entering Manage, with a one-minute cache and a retry action on failure. Save notes and preview reversible Windows user-skill archives and restores in the same view.

  • The separate Setup page has been removed. The companion setup Skill and read-only recording diagnostic tool remain available in chat.

The complete product combines a local MCP service, an activity UI, and two companion Skills for setup and inspection. Installing a Skill alone does not install the service. Skill adoption recording requires caller participation. Plugin calls support opt-in local host-event capture; host events and assistant reports remain separate views. Local logs do not establish complete host-wide coverage. Missing records do not mean a Skill was never used.

Related MCP server: Skill Graph MCP Server

Architecture

Architecture: local evidence feeds the MCP service, which serves the Codex dashboard and keeps separate local records

The service reads Skill metadata and checks installed-plugin state through the Codex CLI. With automatic capture enabled, it incrementally extracts MCP call events from readable local session logs. Explicit adoption and call reports arrive through MCP tools. The service deduplicates records, aggregates them by date and source, and serves the Activity and Manage views through MCP Apps.

Versioned local JSON stores keep activity, coverage, notes and recovery metadata; raw conversations are not copied into the activity ledgers. Health checks are read-only, redundancy results are candidates, and user-Skill archive or restore actions use a preview and explicit authorization. MCP results are returned to the host under its policies; installation discovery may contact configured marketplaces. See data flow and counting rules.

Where it runs

The current package targets local Codex use, including the Codex desktop app. It does not run directly in ChatGPT on the web.

Environment

Current project support

Codex desktop with a local executor

Intended UI host; Windows installation and MCP checks passed. Native sidebar interaction remains incompletely verified.

Codex CLI

Local installation, registration and MCP tools are tested; the CLI does not provide the graphical dashboard.

ChatGPT web / cloud execution

Not supported by this release. No hosted MCP endpoint or configured web bridge is included.

Other desktop hosts / operating systems

Not verified by this project.

The service launches through local stdio using Node.js, reads local Codex files, and uses the Codex CLI for installed-plugin discovery. Signing in to the same account on the web does not connect these local resources. Browser screenshots and UI tests demonstrate the frontend, not a ChatGPT web deployment. Web support would require a supported remote connection or bridge plus explicit handling of local data access, authentication and permissions; changing the manifest alone is insufficient.

Platform references (checked 2026-10-29): desktop-only plugins and MCP host/configuration boundaries. These describe platform capabilities, not certification of this project.

Quick start: complete manager

Tested on Windows. Requires Node.js 20.18+ and the Codex CLI on PATH. This is an independent project, not an official OpenAI product or an official-directory listing.

Download the v0.5.0 complete plugin ZIP, verify its checksum on the release page, extract it to a permanent directory, and open PowerShell in the folder containing install.mjs:

node install.mjs install
node install.mjs check

The check should show the installed 0.5.0 plugin with enabled set to true and a matching registered source. Reopen Codex, then ask it to open Plugin & Skill Manager or run the packaged setup Skill. The complete package includes the service, UI and both companion Skills. Keep the extracted folder in place; no npm install or build is required.

Version-specific regression and installation results are recorded in the validation log. Browser management checks passed; native-desktop interaction remains unverified. See validation, known limitations, and upgrade/uninstall. Developers can still install from source.

Official directory status

The v0.5.0 desktop submission ZIP is available separately on the release page. It contains the manifest and local MCP server at its root; use this for portal upload, not the installation ZIP. As of 2026-10-27, a 0.3.1 draft has been uploaded, but it has not been submitted for review, approved, or published. At that check, the portal reported that it was not accepting more skills-only submissions. Recognition and eligibility for the bundled local MCP server remain unresolved. The GitHub release is available for manual installation; a saved portal draft is not an official listing.

Optional: companion Skill

For third-party Skill directories, see the dependency-aware installation guide.

The complete plugin already includes these Skills. For the alternative source/MCP route, install the companion instructions from your downloaded source into your chosen project. The pinned skills CLI requires Node.js 22.20+, independently of the manager's Node requirement:

# Run in the project where you want to use the Skill.
# Replace this path with your extracted source directory.
npx skills@1.7.4 add "C:\tools\plugin-skill-manager" --skill skill-manager --agent codex --copy

This local-source route was tested in an isolated project. On November 1, 2026, one remote GitHub installation through the official CLI succeeded using the existing local proxy, resolving the earlier clone failure for that check. The skills.sh detail page was available; search indexing remained unconfirmed. See verification notes and companion Skill installation.

Start with: “Open Plugin & Skill Manager and explain the difference between adoption and call counts.” To collect new records, follow adoption recording or plugin-call recording. The companion Skill alone does not enable background collection.

Uninstall the complete plugin, preserving its local data:

node install.mjs uninstall

Environment health

The v0.5.0 source and complete plugin download share the same version. See the v0.5.0 integration notes. Includes Environment health, grouped redundancy reports, installation-unit boundaries and packaged assessment principles. Functional-candidate retrieval still uses lexical/task clues and requires review; shared auxiliary steps alone do not establish redundancy. The screenshot above was captured with v0.4.14. See validation for verified behavior and limits.

Demo and feedback

Current demo and capture details · Chinese README · Report an issue

Useful feedback: installation steps that failed, confusing metric definitions, or missing workflows. Share sanitized errors and reproduction steps; omit private chats, paths and credentials.

Evidence and Counting

Signal

What it means

Skill adoption

The assistant explicitly records an adoption with record_skill_use. Repeated adoptions count separately.

Plugin tool call

An opted-in local host-event collector or an explicit caller records a tool-call attempt. Failed attempts count too.

Historical read request

A separate inferred signal; reading a Skill file can also mean review or audit.

Explicit mention

A separate historical intent signal, not proof of execution.

Adoptions and plugin calls are separate metrics and are never added together. Skill adoptions are caller-reported; plugin calls can be collected from local host events after explicit opt-in. No records means unknown history, not never used. Enabling a recording tool does not backfill previous activity.

Retries reuse the same batch or call ID for deduplication. New adoptions remain separate even within one conversation. Skill counts use file instances; same-name copies retain their source identity. Ambiguous attribution is not distributed across instances.

Daily totals use the full index rather than the limited evidence list. Rankings use the intersection of the selected dates and recorded coverage. With automatic capture disabled, refresh reads existing statistics only. When explicitly enabled, local polling and visible-panel refresh honor the shared collection interval; coverage remains limited to readable local host events.

Privacy

Runtime records are stored locally under $CODEX_HOME/skill-manager/data-v1 (default ~/.codex/skill-manager/data-v1); archives are separately kept under $CODEX_HOME/skill-manager-archive. The application adds no telemetry or upload endpoint, but MCP tool results are returned to the host and may be used in its model context under the host's policies. Plugin-installation checks, including automatic checks on entering Manage, may contact configured marketplaces. See privacy and data flow, project terms, support, and migration.

The repository contains code, documentation and synthetic test fixtures. The current English README image shows explicitly selected, sanitized real aggregate activity from the maintainer. It contains dates and counts, not chat text, tool arguments, session identifiers, credentials or local file paths. Historical demo images remain labeled as synthetic.

Development

See architecture, configuration and dependency management. pnpm run verify runs contract, architecture and type checks plus the full test suite.

pnpm test builds the UI and runs unit and stdio MCP integration tests with synthetic temporary directories. UI resources use content-hashed URLs; retained revisions let open panels request their matching version. Do not remove revisions while older panels remain open.

The frontend uses the MCP Apps SDK; backend tools use MCP SDK v1. Dependency versions are pinned. No Python runtime is required.

pnpm run build:marketplace now creates a self-contained development catalog with the runtime package and companion Skill. See marketplace setup and the resolved Windows installation issue. This is not yet a verified remote GitHub marketplace or published npm package. Protocol and synthetic UI tests do not establish native-sidebar acceptance.

License

Copyright 2026 Zishen Wang. Licensed under Apache-2.0; see NOTICE. Dependencies retain their own licenses. Generated standalone packages include third-party license texts.

Available Tools

13 tools
cancel_skill_plan取消管理预览A

使尚未执行的归档或恢复预览失效,不修改技能文件。

Inputs: token (required).

check_plugin_status核对插件安装状态A

通过官方 app-server plugin/installed 读取已安装插件版本和启用状态,不拉取远程市场目录,不扫描活动日志。

Inputs: no named parameters.

check_recording_status检查记录接入状态A
Read-only

只读检查已有采用和插件调用账本,返回总数与最近接收时间。不写测试事件、不读取聊天、不证明其他会话已经接入;空账本不等于没有实际活动。

Inputs: no named parameters.

configure_auto_capture配置本机全局插件调用采集A

按用户要求启停跨会话本地日志采集并设置 5 至 3600 秒检查间隔。启用后读取 sessions 和 archived_sessions 的真实 MCP 调用事件;仅保存标识与时间。配置在本机持久化,不调用模型、不联网、不自动确认 Skill 采用。

Inputs: enabled (required), intervalSeconds.

execute_skill_plan执行已预览的归档或恢复A
Destructive

仅在用户已针对该具体预览明确要求执行后调用。按一次性预览令牌移动用户技能目录,保留恢复记录;系统和插件只读。

Inputs: token (required).

import_usage增量导入本机使用信号A

读取本机 Codex sessions 和 archived_sessions 日志,增量提取显式标记与读取请求。只将信号、来源定位和读取位置写入独立持久数据目录中的 usage-index.json,不保存聊天正文,不联网。

Inputs: no named parameters.

open_skill_managerPlugin & Skill ManagerA

扫描技能并打开管理器。启用自动采集时增量读取本机日志的宿主工具事件,只保存调用标识和时间,不保存正文。

Inputs: no named parameters.

preview_skill_archive预览用户技能归档A
Read-only

只读预览用户技能目录的归档范围与影响,不执行移动。系统和插件内容不可归档。

Inputs: skillId (required).

preview_skill_restore预览用户技能恢复B
Read-only

只读核对已归档技能并生成恢复预览;原位置已有内容时拒绝覆盖。

Inputs: archiveId (required).

query_usage按期间查询使用信号A

查询已有 Skill 索引与插件账本;启用自动采集时先增量接收本机宿主调用事件。时间需带时区,包含起点、不包含终点。

Inputs: to, from, timeZone.

record_plugin_calls记录插件工具调用A

实际调用插件工具后就近合批记录,每项一次调用尝试(失败也计);id 使用调用 ID 或事先生成的 UUID,记录重试沿用,新的实际调用使用新 ID。plugin 为工具所属插件命名空间,如 github;tool 为完整 mcp__codex_apps__github_get_repo 名称。仅记录已发生调用,不记录发现、计划或历史补记。不传参数或结果正文,与 Skill 采用分开。

Inputs: calls (required).

record_skill_use记录技能采用A

采用技能时批量记录。id 为新批次唯一 ID(重试沿用);skills 为技能名、插件名:技能名或准确路径,可重复,每项计一次。不用于历史补记或仅阅读审计。

Inputs: id (required), skills (required).

save_skill_note保存技能保留备注A

保存用户提供的本地备注和保留标记,不修改技能文件或启停状态。需基于当前文件指纹和备注版本,冲突时拒绝覆盖。

Inputs: id (required), note (required), retain (required), fingerprint (required), +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 13 tools

    Disambiguation3/5

    Tools split into recognizable areas (capture config, log import/query, recording, notes, archive/restore), but boundaries between import_usage, open_skill_manager, query_usage, and check_recording_status are subtle because several involve local log/ledger ingestion or inspection. Descriptions help, but an agent could hesitate between querying existing data vs. incrementally importing more host events.

    Naming Consistency5/5

    All tool names use snake_case and a consistent verb-first pattern such as configure_auto_capture, check_plugin_status, record_plugin_calls, preview_skill_archive, and cancel_skill_plan. There is no disruptive mixing of camelCase or vague standalone verbs.

    Tool Count5/5

    Thirteen tools is within a reasonable range for a plugin/skill management surface that includes capture configuration, usage import/query, recording, notes, and archive/restore lifecycle operations. Each tool maps to a concrete action or safety boundary.

    Completeness3/5

    The surface covers usage ingestion/query, call recording, skill notes, plugin status checks, and skill archive/restore previews and execution. However, plugin lifecycle operations such as install/update/enable/disable are absent, and direct skill enable/disable or richer skill management operations are not exposed.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Enables AI agents to search and inspect a Windows machine's local inventory of agents, skills, MCP servers, projects, and cross-agent work records, with tools for full-text search, overviews, project details, and reindexing. All data stays local and only matching snippets enter model context.
      8
      3
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables an AI assistant to invoke local file operations, multi-engine web search, weather, system commands, developer utilities, and code execution over MCP, and includes a web dashboard for managing endpoints, testing tools, and viewing live logs.
      43 npm
      1
      MIT